medrs 0.1.2

Ultra-high-performance medical imaging I/O for deep learning
Documentation
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
//! High-performance NIfTI I/O operations.
//!
//! Optimizations:
//! - Memory-mapped reading for uncompressed files
//! - libdeflate-based decompression for .nii.gz

use super::header::NiftiHeader;
use super::image::NiftiImage;
use crate::error::{Error, Result};
use flate2::bufread::GzDecoder;
use libdeflater::Decompressor;
use memmap2::Mmap;
use rand::Rng;
use std::collections::HashMap;
use std::fs::File;
use std::io::{BufReader, BufWriter, Cursor, Read, Write};
use std::path::{Path, PathBuf};
use std::sync::Arc;

use crate::transforms::{self, Interpolation};

/// Load a NIfTI image from file.
///
/// Supports both `.nii` and `.nii.gz` formats with automatic detection.
///
/// # Example
/// ```ignore
/// let img = medrs::nifti::load("brain.nii.gz")?;
/// let data = img.to_f32();
/// ```
#[must_use = "this function returns a loaded image that should be used"]
pub fn load<P: AsRef<Path>>(path: P) -> Result<NiftiImage> {
    let path = path.as_ref();
    let is_gzipped = path.extension().is_some_and(|e| e == "gz");

    if is_gzipped {
        load_gzipped(path)
    } else {
        load_uncompressed(path)
    }
}

/// Load uncompressed .nii file using memory mapping for speed.
#[allow(unsafe_code)]
fn load_uncompressed(path: &Path) -> Result<NiftiImage> {
    let file = File::open(path)?;
    // SAFETY: Memory mapping is safe because:
    // 1. The file was just opened successfully
    // 2. The mmap is read-only and won't be modified
    // 3. If the file is modified externally, data may become inconsistent but no UB
    let mmap = unsafe { Mmap::map(&file)? };

    let header = NiftiHeader::from_bytes(&mmap)?;
    let offset = header.vox_offset as usize;
    let data_size = header.data_size();

    if mmap.len() < offset + data_size {
        return Err(Error::Io(std::io::Error::new(
            std::io::ErrorKind::UnexpectedEof,
            "file truncated",
        )));
    }

    let arc = Arc::new(mmap);
    Ok(NiftiImage::from_shared_mmap(header, arc, offset, data_size))
}

/// Load gzipped .nii.gz file with streaming decompression.
fn load_gzipped(path: &Path) -> Result<NiftiImage> {
    // Read compressed bytes
    let compressed = std::fs::read(path)?;

    // First, decompress only the header to discover expected size
    let mut header_buf = [0u8; NiftiHeader::SIZE];
    {
        let cursor = Cursor::new(&compressed);
        let mut decoder = GzDecoder::new(cursor);
        decoder.read_exact(&mut header_buf)?;
    }
    let header = NiftiHeader::from_bytes(&header_buf)?;

    let offset = header.vox_offset as usize;
    let data_size = header.data_size();
    let total_size = offset + data_size;

    // Decompress entire payload in one go using libdeflate (fast single-thread)
    let mut output = vec![0u8; total_size];
    let mut decompressor = Decompressor::new();
    let written = decompressor
        .gzip_decompress(&compressed, &mut output)
        .map_err(|e| Error::Decompression(format!("{}", e)))?;

    if written != total_size {
        return Err(Error::Decompression(format!(
            "decompressed size {} did not match expected {}",
            written, total_size
        )));
    }

    let bytes = Arc::new(output);
    Ok(NiftiImage::from_shared_bytes(
        header, bytes, offset, data_size,
    ))
}

/// Save a NIfTI image to file.
///
/// Format is determined by file extension (`.nii` or `.nii.gz`).
///
/// # Example
/// ```ignore
/// medrs::nifti::save(&img, "output.nii.gz")?;
/// ```
pub fn save<P: AsRef<Path>>(image: &NiftiImage, path: P) -> Result<()> {
    image.header().validate()?;

    let path = path.as_ref();
    let is_gzipped = path.extension().is_some_and(|e| e == "gz");

    if is_gzipped {
        save_gzipped(image, path)
    } else {
        save_uncompressed(image, path)
    }
}

fn save_uncompressed(image: &NiftiImage, path: &Path) -> Result<()> {
    let file = File::create(path)?;
    let mut writer = BufWriter::with_capacity(1024 * 1024, file);

    // Write header
    let header_bytes = image.header().to_bytes();
    writer.write_all(&header_bytes)?;

    // Padding to vox_offset (typically 352)
    let padding = image.header().vox_offset as usize - NiftiHeader::SIZE;
    if padding > 0 {
        writer.write_all(&vec![0u8; padding])?;
    }

    // Write data
    let data = image.data_to_bytes()?;
    writer.write_all(&data)?;
    writer.flush()?;

    Ok(())
}

fn save_gzipped(image: &NiftiImage, path: &Path) -> Result<()> {
    // Build uncompressed payload in memory first
    let header_bytes = image.header().to_bytes();
    let padding = image.header().vox_offset as usize - NiftiHeader::SIZE;
    let data = image.data_to_bytes()?;

    // Assemble full uncompressed payload
    let total_size = header_bytes.len() + padding + data.len();
    let mut uncompressed = Vec::with_capacity(total_size);
    uncompressed.extend_from_slice(&header_bytes);
    uncompressed.resize(uncompressed.len() + padding, 0u8);
    uncompressed.extend_from_slice(&data);

    // Use libdeflate for fast single-shot compression (same as we use for decompression)
    // Level 1 = fastest, good balance of speed vs compression ratio
    let mut compressor = libdeflater::Compressor::new(libdeflater::CompressionLvl::fastest());

    // Allocate output buffer (worst case: slightly larger than input for incompressible data)
    let max_compressed_size = compressor.gzip_compress_bound(uncompressed.len());
    let mut compressed = vec![0u8; max_compressed_size];

    let actual_size = compressor
        .gzip_compress(&uncompressed, &mut compressed)
        .map_err(|e| Error::Io(std::io::Error::other(format!("compression failed: {e:?}"))))?;

    compressed.truncate(actual_size);

    // Write compressed data to file
    let mut file = File::create(path)?;
    file.write_all(&compressed)?;

    Ok(())
}

/// Load only the header from a NIfTI file (fast metadata inspection).
#[allow(unsafe_code)]
pub fn load_header<P: AsRef<Path>>(path: P) -> Result<NiftiHeader> {
    let path = path.as_ref();
    let is_gzipped = path.extension().is_some_and(|e| e == "gz");

    if is_gzipped {
        let file = File::open(path)?;
        let buf_reader = BufReader::new(file);
        let mut decoder = GzDecoder::new(buf_reader);
        let mut header_buf = vec![0u8; NiftiHeader::SIZE];
        decoder.read_exact(&mut header_buf)?;
        NiftiHeader::from_bytes(&header_buf)
    } else {
        let file = File::open(path)?;
        // SAFETY: Memory mapping is safe - file just opened, read-only access
        let mmap = unsafe { Mmap::map(&file)? };
        NiftiHeader::from_bytes(&mmap)
    }
}

/// Configuration for loading a cropped region with optional transforms.
pub struct LoadCroppedConfig {
    /// Desired output shape after all transforms [d, h, w]
    pub output_shape: [usize; 3],
    /// Target voxel spacing [mm] for resampling (None = keep original)
    pub target_spacing: Option<[f32; 3]>,
    /// Target orientation (None = keep original)
    pub target_orientation: Option<crate::transforms::Orientation>,
    /// Optional offset in output space for non-centered crops [d, h, w]
    pub output_offset: Option<[usize; 3]>,
}

impl Default for LoadCroppedConfig {
    fn default() -> Self {
        Self {
            output_shape: [64, 64, 64],
            target_spacing: None,
            target_orientation: None,
            output_offset: None,
        }
    }
}

/// Load a cropped region from a NIfTI file without loading the entire volume.
///
/// This is extremely efficient for training pipelines that load large volumes
/// just to crop small patches (e.g., loading 64^3 patch from 256^3 volume).
///
/// Advanced version that can compute minimal raw region needed to achieve
/// desired output after optional reorientation and resampling.
///
/// Arguments:
///   path: Path to NIfTI file (must be uncompressed .nii)
///   config: Configuration specifying desired output and optional transforms
///
/// Returns: NiftiImage with cropped data (as owned array)
#[must_use = "this function returns a loaded image that should be used"]
#[allow(unsafe_code)]
pub fn load_cropped_config<P: AsRef<Path>>(
    path: P,
    config: LoadCroppedConfig,
) -> Result<NiftiImage> {
    let path = path.as_ref();

    ensure_uncompressed(path)?;

    let file = File::open(path)?;
    // SAFETY: Memory mapping is safe - file just opened, read-only access
    let mmap = unsafe { Mmap::map(&file)? };
    let header = NiftiHeader::from_bytes(&mmap)?;
    let data_offset = header.vox_offset as usize;
    let shape = header.shape();
    let crop_offset = config
        .output_offset
        .unwrap_or(compute_center_offset(&shape, &config.output_shape)?);

    let cropped = copy_cropped_region(
        &header,
        &mmap,
        data_offset,
        crop_offset,
        config.output_shape,
    )?;

    let mut output = cropped;

    if let Some(target_orient) = config.target_orientation {
        output = transforms::reorient(&output, target_orient)?;
    }

    if let Some(target_spacing) = config.target_spacing {
        output =
            transforms::resample_to_spacing(&output, target_spacing, Interpolation::Trilinear)?;
    }

    Ok(output)
}

/// Simple version of load_cropped.
#[must_use = "this function returns a loaded image that should be used"]
#[allow(unsafe_code)]
pub fn load_cropped<P: AsRef<Path>>(
    path: P,
    crop_offset: [usize; 3],
    crop_shape: [usize; 3],
) -> Result<NiftiImage> {
    let path = path.as_ref();

    ensure_uncompressed(path)?;

    let file = File::open(path)?;
    // SAFETY: Memory mapping is safe - file just opened, read-only access
    let mmap = unsafe { Mmap::map(&file)? };
    let header = NiftiHeader::from_bytes(&mmap)?;
    let data_offset = header.vox_offset as usize;

    copy_cropped_region(&header, &mmap, data_offset, crop_offset, crop_shape)
}

fn ensure_uncompressed(path: &Path) -> Result<()> {
    if path.extension().is_some_and(|e| e == "gz") {
        return Err(Error::InvalidDimensions(
            "load_cropped only supports uncompressed .nii files".to_string(),
        ));
    }
    Ok(())
}

fn compute_center_offset(full_shape: &[u16], crop_shape: &[usize; 3]) -> Result<[usize; 3]> {
    if full_shape.len() < 3 {
        return Err(Error::InvalidDimensions(
            "expected at least 3 spatial dimensions".to_string(),
        ));
    }

    Ok([
        (full_shape[0] as usize).saturating_sub(crop_shape[0]) / 2,
        (full_shape[1] as usize).saturating_sub(crop_shape[1]) / 2,
        (full_shape[2] as usize).saturating_sub(crop_shape[2]) / 2,
    ])
}

fn copy_cropped_region(
    header: &NiftiHeader,
    mmap: &Mmap,
    data_offset: usize,
    crop_offset: [usize; 3],
    crop_shape: [usize; 3],
) -> Result<NiftiImage> {
    let full_shape = header.shape();
    if full_shape.len() < 3 {
        return Err(Error::InvalidDimensions(
            "expected at least 3 spatial dimensions".to_string(),
        ));
    }

    // Validate crop_shape has no zero dimensions
    for (i, &dim) in crop_shape.iter().enumerate() {
        if dim == 0 {
            return Err(Error::InvalidDimensions(format!(
                "Crop shape dimension {} cannot be zero",
                i
            )));
        }
    }

    for i in 0..3 {
        if crop_offset[i] + crop_shape[i] > full_shape[i] as usize {
            return Err(Error::InvalidDimensions(format!(
                "Crop region exceeds dimension {}: {} + {} > {}",
                i, crop_offset[i], crop_shape[i], full_shape[i]
            )));
        }
    }

    let elem_size = header.datatype.size();
    let dim0 = full_shape[0] as usize; // First dimension (fastest changing in F-order)
    let dim1 = full_shape.get(1).copied().unwrap_or(1) as usize;

    // Calculate total_bytes with overflow checking
    let total_bytes = crop_shape[0]
        .checked_mul(crop_shape[1])
        .and_then(|v| v.checked_mul(crop_shape[2]))
        .and_then(|v| v.checked_mul(elem_size))
        .ok_or_else(|| {
            Error::InvalidDimensions(format!(
                "Crop region too large: {:?} x {} bytes would overflow",
                crop_shape, elem_size
            ))
        })?;
    let mut buffer = vec![0u8; total_bytes];

    // Calculate expected_data_size with overflow checking
    let expected_data_size = (full_shape[0] as usize)
        .checked_mul(full_shape[1] as usize)
        .and_then(|v| v.checked_mul(full_shape[2] as usize))
        .and_then(|v| v.checked_mul(elem_size))
        .ok_or_else(|| {
            Error::InvalidDimensions(format!(
                "Volume too large: {:?} x {} bytes would overflow",
                &full_shape[..3],
                elem_size
            ))
        })?;

    let total_required = data_offset
        .checked_add(expected_data_size)
        .ok_or_else(|| Error::InvalidDimensions("Data offset + size would overflow".to_string()))?;

    if mmap.len() < total_required {
        return Err(Error::InvalidDimensions(format!(
            "File too small: need {} bytes for data but mmap is {} bytes (offset {})",
            expected_data_size,
            mmap.len(),
            data_offset
        )));
    }

    // NIfTI uses F-order (column-major): first index changes fastest
    // F-order linear index = x + y * dim0 + z * dim0 * dim1
    let row_bytes = crop_shape[0] * elem_size;
    let mmap_slice = mmap.as_ref();

    // For larger crops, use parallel copying across z-slices
    // Threshold: ~64KB of data per slice makes parallelization worthwhile
    let slice_bytes = crop_shape[0] * crop_shape[1] * elem_size;
    let use_parallel = slice_bytes > 65536 && crop_shape[2] >= 4;

    if use_parallel {
        use rayon::prelude::*;

        // Split buffer into z-slices and copy in parallel
        let slices_per_z = crop_shape[1];
        let bytes_per_z = slices_per_z * row_bytes;

        buffer
            .par_chunks_mut(bytes_per_z)
            .enumerate()
            .for_each(|(z, z_buffer)| {
                let src_z = crop_offset[2] + z;
                let z_offset = src_z * dim0 * dim1;

                for y in 0..crop_shape[1] {
                    let src_y = crop_offset[1] + y;
                    let src_index = crop_offset[0] + src_y * dim0 + z_offset;
                    let src_byte = data_offset + src_index * elem_size;

                    let dst_start = y * row_bytes;
                    z_buffer[dst_start..dst_start + row_bytes]
                        .copy_from_slice(&mmap_slice[src_byte..src_byte + row_bytes]);
                }
            });
    } else {
        // Sequential copy for small crops (lower overhead)
        let mut dst_cursor = 0;
        for z in 0..crop_shape[2] {
            let src_z = crop_offset[2] + z;
            let z_offset = src_z * dim0 * dim1;

            for y in 0..crop_shape[1] {
                let src_y = crop_offset[1] + y;
                let src_index = crop_offset[0] + src_y * dim0 + z_offset;
                let src_byte = data_offset + src_index * elem_size;

                let src_range = src_byte..src_byte + row_bytes;
                let dst_range = dst_cursor..dst_cursor + row_bytes;
                buffer[dst_range].copy_from_slice(&mmap_slice[src_range]);
                dst_cursor += row_bytes;
            }
        }
    }

    let mut new_header = header.clone();
    new_header.ndim = 3;
    new_header.dim = [1u16; 7];
    for (i, &s) in crop_shape.iter().enumerate() {
        new_header.dim[i] = s as u16;
    }
    new_header.vox_offset = NiftiHeader::default().vox_offset;

    // Translate affine by crop offset
    let mut affine = new_header.affine();
    for row in affine.iter_mut().take(3) {
        row[3] += row[0] * crop_offset[0] as f32
            + row[1] * crop_offset[1] as f32
            + row[2] * crop_offset[2] as f32;
    }
    new_header.set_affine(affine);

    let data_len = buffer.len();
    Ok(NiftiImage::from_shared_bytes(
        new_header,
        Arc::new(buffer),
        0,
        data_len,
    ))
}

/// Configuration for streaming crop loader optimized for training pipelines.
pub struct CropLoaderConfig {
    /// Patch size to extract [d, h, w]
    pub patch_size: [usize; 3],
    /// Number of patches per volume
    pub patches_per_volume: usize,
    /// Overlap between patches [d, h, w] in voxels
    pub patch_overlap: [usize; 3],
    /// Whether to randomize patch positions
    pub randomize: bool,
}

impl Default for CropLoaderConfig {
    fn default() -> Self {
        Self {
            patch_size: [64, 64, 64],
            patches_per_volume: 4,
            patch_overlap: [0, 0, 0],
            randomize: false,
        }
    }
}

/// Streaming crop loader that efficiently extracts multiple patches from volumes.
///
/// This maintains memory efficiency while extracting multiple patches per volume,
/// perfect for training pipelines.
pub struct CropLoader {
    volumes: Vec<PathBuf>,
    current_volume: usize,
    patches_extracted: usize,
    config: CropLoaderConfig,
}

impl CropLoader {
    /// Create a new crop loader for the given volumes.
    pub fn new<P: AsRef<Path>>(volumes: Vec<P>, config: CropLoaderConfig) -> Self {
        Self {
            volumes: volumes.iter().map(|p| p.as_ref().to_path_buf()).collect(),
            current_volume: 0,
            patches_extracted: 0,
            config,
        }
    }

    /// Extract the next patch from the training set.
    #[allow(unsafe_code)]
    pub fn next_patch(&mut self) -> Result<NiftiImage> {
        if self.current_volume >= self.volumes.len() {
            return Err(Error::Exhausted("all volumes processed".to_string()));
        }

        if self.config.patches_per_volume == 0 {
            return Err(Error::InvalidDimensions(
                "patches_per_volume must be positive".to_string(),
            ));
        }

        // Load current volume header to get dimensions
        let path = &self.volumes[self.current_volume];
        let file = File::open(path)?;
        // SAFETY: Memory mapping is safe - file just opened, read-only access
        let mmap = unsafe { Mmap::map(&file)? };
        let header = NiftiHeader::from_bytes(&mmap)?;
        let volume_shape = header.shape();

        if volume_shape.len() < 3 {
            return Err(Error::InvalidDimensions(
                "expected at least 3 spatial dimensions".to_string(),
            ));
        }

        for (i, (&dim, &patch_dim)) in volume_shape
            .iter()
            .zip(self.config.patch_size.iter())
            .take(3)
            .enumerate()
        {
            if patch_dim == 0 {
                return Err(Error::InvalidDimensions(format!(
                    "patch_size[{}] must be positive",
                    i
                )));
            }
            if patch_dim > dim as usize {
                return Err(Error::InvalidDimensions(format!(
                    "patch_size[{}]={} cannot exceed image dimension[{}]={}",
                    i, patch_dim, i, dim
                )));
            }
        }

        for i in 0..3 {
            if self.config.patch_overlap[i] >= self.config.patch_size[i] {
                return Err(Error::InvalidDimensions(
                    "patch_overlap must be smaller than patch_size in all dimensions".to_string(),
                ));
            }
        }

        // Calculate patch positions
        let patch_positions = if self.config.randomize {
            self.random_patch_positions(&volume_shape)
        } else {
            self.grid_patch_positions(&volume_shape)?
        };

        // Get the next patch position
        if self.patches_extracted >= patch_positions.len() {
            // Move to next volume
            self.current_volume += 1;
            self.patches_extracted = 0;
            return self.next_patch(); // Recursive call for next volume
        }

        let patch_offset = patch_positions[self.patches_extracted];
        self.patches_extracted += 1;

        // Use the efficient load_cropped function
        load_cropped(path, patch_offset, self.config.patch_size)
    }

    /// Calculate grid-based patch positions.
    fn grid_patch_positions(&self, shape: &[u16]) -> Result<Vec<[usize; 3]>> {
        let [pd, ph, pw] = self.config.patch_size;
        let [od, oh, ow] = self.config.patch_overlap;

        let step_d = pd.saturating_sub(od);
        let step_h = ph.saturating_sub(oh);
        let step_w = pw.saturating_sub(ow);

        if step_d == 0 || step_h == 0 || step_w == 0 {
            return Err(Error::InvalidDimensions(
                "patch_size must be larger than patch_overlap in all dimensions".to_string(),
            ));
        }

        let mut positions = Vec::new();
        // Use saturating_sub to prevent underflow when patch is larger than volume
        let max_d = (shape[0] as usize).saturating_sub(pd);
        let max_h = (*shape.get(1).unwrap_or(&1) as usize).saturating_sub(ph);
        let max_w = (*shape.get(2).unwrap_or(&1) as usize).saturating_sub(pw);

        for d in (0..=max_d).step_by(step_d) {
            for h in (0..=max_h).step_by(step_h) {
                for w in (0..=max_w).step_by(step_w) {
                    positions.push([d, h, w]);
                }
            }
        }

        Ok(positions)
    }

    /// Calculate random patch positions.
    fn random_patch_positions(&self, shape: &[u16]) -> Vec<[usize; 3]> {
        use rand::thread_rng;

        // Use saturating_sub to prevent underflow when patch is larger than volume
        let max_d =
            (*shape.first().unwrap_or(&1) as usize).saturating_sub(self.config.patch_size[0]);
        let max_h =
            (*shape.get(1).unwrap_or(&1) as usize).saturating_sub(self.config.patch_size[1]);
        let max_w =
            (*shape.get(2).unwrap_or(&1) as usize).saturating_sub(self.config.patch_size[2]);

        let mut rng = thread_rng();
        let mut positions = Vec::new();

        for _ in 0..self.config.patches_per_volume {
            positions.push([
                rng.gen_range(0..=max_d),
                rng.gen_range(0..=max_h),
                rng.gen_range(0..=max_w),
            ]);
        }

        positions
    }
}

/// Batch loader for efficient training pipeline data loading.
///
/// This combines multiple volumes and patch extraction into a memory-efficient
/// streaming interface optimized for high-throughput training.
pub struct BatchLoader {
    loader: CropLoader,
    batch_size: usize,
}

impl BatchLoader {
    /// Create a new batch loader.
    pub fn new<P: AsRef<Path>>(volumes: Vec<P>, batch_size: usize) -> Self {
        let config = CropLoaderConfig::default();
        Self {
            loader: CropLoader::new(volumes, config),
            batch_size,
        }
    }

    /// Load the next batch of patches.
    pub fn next_batch(&mut self) -> Result<Vec<NiftiImage>> {
        if self.batch_size == 0 {
            return Err(Error::InvalidDimensions(
                "batch_size must be positive".to_string(),
            ));
        }

        let mut batch = Vec::with_capacity(self.batch_size);

        for _ in 0..self.batch_size {
            match self.loader.next_patch() {
                Ok(patch) => batch.push(patch),
                Err(Error::Exhausted(_)) => break,
                Err(e) => return Err(e),
            }
        }

        if batch.is_empty() {
            Err(Error::Exhausted("no more patches available".to_string()))
        } else {
            Ok(batch)
        }
    }
}

/// High-performance prefetch and caching system for training pipelines.
///
/// Maintains an LRU cache of loaded patches and prefetches upcoming data
/// to maximize I/O throughput while iterating over many volumes.
pub struct TrainingDataLoader {
    /// Volume file paths
    volumes: Vec<PathBuf>,
    /// Patch configuration
    config: CropLoaderConfig,
    /// LRU cache of loaded patches (volume_index -> cached patches)
    cache: HashMap<usize, Vec<NiftiImage>>,
    /// Maximum cache size in patches
    max_cache_size: usize,
    /// Current volume index
    current_volume: usize,
    /// Current patch index within volume
    current_patch: usize,
    /// Prefetch queue
    prefetch_queue: Vec<(usize, Vec<[usize; 3]>)>,
    /// Total patches processed
    patches_processed: usize,
}

impl TrainingDataLoader {
    /// Create a new training data loader with prefetching.
    ///
    /// Args:
    ///     volumes: List of NIfTI file paths
    ///     config: Patch extraction configuration
    ///     cache_size: Maximum number of patches to cache (default: 1000)
    ///
    /// Example:
    ///     ```rust
    ///     let loader = TrainingDataLoader::new(
    ///         vec!["vol1.nii", "vol2.nii"],
    ///         CropLoaderConfig {
    ///             patch_size: [64, 64, 64],
    ///             patches_per_volume: 4,
    ///             patch_overlap: [0, 0, 0],
    ///             randomize: true,
    ///         },
    ///         1000, // Cache 1000 patches
    ///     );
    ///     ```
    pub fn new<P: AsRef<Path>>(
        volumes: Vec<P>,
        config: CropLoaderConfig,
        cache_size: usize,
    ) -> Result<Self> {
        let volumes: Vec<PathBuf> = volumes
            .into_iter()
            .map(|p| p.as_ref().to_path_buf())
            .collect();

        if volumes.is_empty() {
            return Err(Error::InvalidDimensions("No volumes provided".to_string()));
        }

        for i in 0..3 {
            if config.patch_size[i] == 0 {
                return Err(Error::InvalidDimensions(format!(
                    "patch_size[{}] must be positive",
                    i
                )));
            }
        }

        if config.patches_per_volume == 0 {
            return Err(Error::InvalidDimensions(
                "patches_per_volume must be positive".to_string(),
            ));
        }

        for i in 0..3 {
            if config.patch_overlap[i] >= config.patch_size[i] {
                return Err(Error::InvalidDimensions(
                    "patch_overlap must be smaller than patch_size in all dimensions".to_string(),
                ));
            }
        }

        let mut loader = Self {
            cache: HashMap::new(),
            max_cache_size: cache_size,
            current_volume: 0,
            current_patch: 0,
            prefetch_queue: Vec::new(),
            patches_processed: 0,
            volumes,
            config,
        };

        // Initialize prefetch queue for first few volumes
        loader.initialize_prefetch()?;
        Ok(loader)
    }

    /// Initialize prefetch queue with upcoming volumes.
    fn initialize_prefetch(&mut self) -> Result<()> {
        // Prefetch first few volumes to fill cache
        let prefetch_count = (self.max_cache_size / self.config.patches_per_volume).min(3);

        for i in 0..prefetch_count.min(self.volumes.len()) {
            let patch_positions = self.compute_patch_positions(&self.volumes[i])?;
            self.prefetch_queue.push((i, patch_positions));
        }

        Ok(())
    }

    /// Get next training patch with automatic prefetching.
    ///
    /// This method maintains a background cache and prefetches upcoming data
    /// to ensure patches are always available with minimal latency.
    ///
    /// Returns: Next training patch
    pub fn next_patch(&mut self) -> Result<NiftiImage> {
        // Check if we need to load current volume's patches
        if !self.cache.contains_key(&self.current_volume) {
            self.load_volume_patches(self.current_volume)?;
        }

        // Get patch from cache (invariant: load_volume_patches ensures key exists)
        let patches = self.cache.get_mut(&self.current_volume).ok_or_else(|| {
            Error::InvalidDimensions("cache invariant violated: volume should be loaded".into())
        })?;

        if self.current_patch >= patches.len() {
            // Move to next volume
            self.current_volume += 1;
            self.current_patch = 0;

            if self.current_volume >= self.volumes.len() {
                return Err(Error::Exhausted("all patches processed".to_string()));
            }

            return self.next_patch();
        }

        let patch = patches.swap_remove(self.current_patch); // Remove for memory efficiency
        self.patches_processed += 1;

        // Trigger prefetch for upcoming volumes if cache is getting low
        if self.cache.len() < 3 && self.current_volume + 2 < self.volumes.len() {
            self.trigger_prefetch(self.current_volume + 2)?;
        }

        Ok(patch)
    }

    /// Load all patches for a volume into cache.
    fn load_volume_patches(&mut self, volume_idx: usize) -> Result<()> {
        let volume_path = &self.volumes[volume_idx];
        let patch_positions = self.compute_patch_positions(volume_path)?;

        let mut patches = Vec::with_capacity(patch_positions.len());

        for position in patch_positions {
            let patch = load_cropped(volume_path, position, self.config.patch_size)?;
            patches.push(patch);
        }

        self.cache.insert(volume_idx, patches);
        Ok(())
    }

    /// Compute patch positions for a volume.
    fn compute_patch_positions(&self, volume_path: &Path) -> Result<Vec<[usize; 3]>> {
        let header = load_header(volume_path)?;
        let shape = header.shape();

        if shape.len() < 3 {
            return Err(Error::InvalidDimensions(
                "expected at least 3 spatial dimensions".to_string(),
            ));
        }

        for (i, (&dim, &patch_dim)) in shape
            .iter()
            .zip(self.config.patch_size.iter())
            .take(3)
            .enumerate()
        {
            if patch_dim > dim as usize {
                return Err(Error::InvalidDimensions(format!(
                    "patch_size[{}]={} cannot exceed image dimension[{}]={}",
                    i, patch_dim, i, dim
                )));
            }
        }

        if self.config.randomize {
            Ok(self.random_patch_positions(shape))
        } else {
            self.grid_patch_positions(shape)
        }
    }

    /// Generate grid-based patch positions.
    fn grid_patch_positions(&self, shape: &[u16]) -> Result<Vec<[usize; 3]>> {
        let [pd, ph, pw] = self.config.patch_size;
        let [od, oh, ow] = self.config.patch_overlap;

        // Use saturating_sub to prevent underflow when patch is larger than volume
        let max_d = (shape[0] as usize).saturating_sub(pd);
        let max_h = (*shape.get(1).unwrap_or(&1) as usize).saturating_sub(ph);
        let max_w = (*shape.get(2).unwrap_or(&1) as usize).saturating_sub(pw);

        let step_d = pd.saturating_sub(od);
        let step_h = ph.saturating_sub(oh);
        let step_w = pw.saturating_sub(ow);

        if step_d == 0 || step_h == 0 || step_w == 0 {
            return Err(Error::InvalidDimensions(
                "patch_size must be larger than patch_overlap in all dimensions".to_string(),
            ));
        }

        let mut positions = Vec::new();
        for d in (0..=max_d).step_by(step_d) {
            for h in (0..=max_h).step_by(step_h) {
                for w in (0..=max_w).step_by(step_w) {
                    positions.push([d, h, w]);
                }
            }
        }

        // Ensure we get exactly patches_per_volume
        if positions.len() > self.config.patches_per_volume {
            positions.truncate(self.config.patches_per_volume);
        } else if positions.len() < self.config.patches_per_volume {
            // Add random positions if needed
            let mut rng = rand::thread_rng();
            while positions.len() < self.config.patches_per_volume {
                let d = rng.gen_range(0..=max_d);
                let h = rng.gen_range(0..=max_h);
                let w = rng.gen_range(0..=max_w);
                positions.push([d, h, w]);
            }
        }

        Ok(positions)
    }

    /// Generate random patch positions.
    fn random_patch_positions(&self, shape: &[u16]) -> Vec<[usize; 3]> {
        let [pd, ph, pw] = self.config.patch_size;

        // Use saturating_sub to prevent underflow when patch is larger than volume
        let max_d = (shape[0] as usize).saturating_sub(pd);
        let max_h = (*shape.get(1).unwrap_or(&1) as usize).saturating_sub(ph);
        let max_w = (*shape.get(2).unwrap_or(&1) as usize).saturating_sub(pw);

        let mut rng = rand::thread_rng();
        let mut positions = Vec::with_capacity(self.config.patches_per_volume);

        for _ in 0..self.config.patches_per_volume {
            positions.push([
                rng.gen_range(0..=max_d),
                rng.gen_range(0..=max_h),
                rng.gen_range(0..=max_w),
            ]);
        }

        positions
    }

    /// Trigger prefetch for upcoming volume.
    fn trigger_prefetch(&mut self, volume_idx: usize) -> Result<()> {
        if volume_idx >= self.volumes.len() {
            return Ok(());
        }

        let patch_positions = self.compute_patch_positions(&self.volumes[volume_idx])?;
        self.prefetch_queue.push((volume_idx, patch_positions));

        Ok(())
    }

    /// Get statistics about the loader performance.
    pub fn stats(&self) -> LoaderStats {
        LoaderStats {
            total_volumes: self.volumes.len(),
            current_volume: self.current_volume,
            cached_volumes: self.cache.len(),
            patches_processed: self.patches_processed,
            cache_size: self.cache.values().map(|patches| patches.len()).sum(),
            max_cache_size: self.max_cache_size,
        }
    }

    /// Reset the loader to start from the beginning.
    pub fn reset(&mut self) -> Result<()> {
        self.cache.clear();
        self.current_volume = 0;
        self.current_patch = 0;
        self.patches_processed = 0;
        self.prefetch_queue.clear();
        self.initialize_prefetch()?;
        Ok(())
    }

    /// Total number of volumes configured for this loader.
    pub fn volumes_len(&self) -> usize {
        self.volumes.len()
    }

    /// Number of patches extracted from each volume.
    pub fn patches_per_volume(&self) -> usize {
        self.config.patches_per_volume
    }
}

/// Performance statistics for the training data loader.
#[derive(Debug, Clone)]
pub struct LoaderStats {
    /// Number of volumes managed by the loader.
    pub total_volumes: usize,
    /// Index of the current volume being processed.
    pub current_volume: usize,
    /// Number of volumes currently cached.
    pub cached_volumes: usize,
    /// Total patches produced so far.
    pub patches_processed: usize,
    /// Current cache size (patches).
    pub cache_size: usize,
    /// Maximum allowed cache size (patches).
    pub max_cache_size: usize,
}

impl std::fmt::Display for LoaderStats {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        writeln!(f, "Training Data Loader Statistics:")?;
        writeln!(f, "  Total volumes: {}", self.total_volumes)?;
        writeln!(
            f,
            "  Current volume: {}/{}",
            self.current_volume + 1,
            self.total_volumes
        )?;
        writeln!(f, "  Cached volumes: {}", self.cached_volumes)?;
        writeln!(f, "  Patches processed: {}", self.patches_processed)?;
        writeln!(
            f,
            "  Cache utilization: {}/{} patches",
            self.cache_size, self.max_cache_size
        )?;
        Ok(())
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use ndarray::s;
    use ndarray::ArrayD;
    use ndarray::ShapeBuilder;
    use tempfile::tempdir;

    fn create_f_order_array(data: Vec<f32>, shape: Vec<usize>) -> ArrayD<f32> {
        let c_order = ArrayD::from_shape_vec(shape.clone(), data).unwrap();
        let mut f_order = ArrayD::zeros(ndarray::IxDyn(&shape).f());
        f_order.assign(&c_order);
        f_order
    }

    #[test]
    fn test_roundtrip_uncompressed() {
        let dir = tempdir().unwrap();
        let path = dir.path().join("test.nii");

        // Create test image with F-order
        let data = create_f_order_array((0..1000).map(|i| i as f32).collect(), vec![10, 10, 10]);
        let affine = [
            [1.0, 0.0, 0.0, 0.0],
            [0.0, 1.0, 0.0, 0.0],
            [0.0, 0.0, 1.0, 0.0],
            [0.0, 0.0, 0.0, 1.0],
        ];
        let img = NiftiImage::from_array(data.clone(), affine);

        // Save and reload
        save(&img, &path).unwrap();
        let loaded = load(&path).unwrap();

        assert_eq!(loaded.shape(), &[10, 10, 10]);
        // Compare in memory order
        let loaded_data = loaded.to_f32().unwrap();
        assert_eq!(
            loaded_data.as_slice_memory_order().unwrap(),
            data.as_slice_memory_order().unwrap()
        );
    }

    #[test]
    fn test_roundtrip_gzipped() {
        let dir = tempdir().unwrap();
        let path = dir.path().join("test.nii.gz");

        let data = create_f_order_array((0..1000).map(|i| i as f32).collect(), vec![10, 10, 10]);
        let affine = [
            [2.0, 0.0, 0.0, -10.0],
            [0.0, 2.0, 0.0, -10.0],
            [0.0, 0.0, 2.0, -10.0],
            [0.0, 0.0, 0.0, 1.0],
        ];
        let img = NiftiImage::from_array(data.clone(), affine);

        save(&img, &path).unwrap();
        let loaded = load(&path).unwrap();

        assert_eq!(loaded.shape(), &[10, 10, 10]);
        assert_eq!(loaded.affine(), affine);
        // Compare in memory order
        let loaded_data = loaded.to_f32().unwrap();
        assert_eq!(
            loaded_data.as_slice_memory_order().unwrap(),
            data.as_slice_memory_order().unwrap()
        );
    }

    #[test]
    fn test_load_cropped_byte_exact() {
        let dir = tempdir().unwrap();
        let path = dir.path().join("test.nii");

        // Create a larger test image for cropping with F-order
        let data = create_f_order_array((0..131072).map(|i| i as f32).collect(), vec![64, 64, 32]);
        let affine = [
            [1.0, 0.0, 0.0, 0.0],
            [0.0, 1.0, 0.0, 0.0],
            [0.0, 0.0, 1.0, 0.0],
            [0.0, 0.0, 0.0, 1.0],
        ];
        let img = NiftiImage::from_array(data.clone(), affine);
        save(&img, &path).unwrap();

        // Test byte-exact cropped loading
        let crop_offset = [16, 16, 8];
        let crop_shape = [32, 32, 16];
        let cropped = load_cropped(&path, crop_offset, crop_shape).unwrap();

        assert_eq!(cropped.shape(), &[32, 32, 16]);

        // Verify the cropped data matches the expected region
        let original_slice = data.slice(s![16..48, 16..48, 8..24]).to_owned();
        let cropped_data = cropped.to_f32().unwrap();

        // Compare by iterating over logical coordinates
        for x in 0..32 {
            for y in 0..32 {
                for z in 0..16 {
                    let expected = original_slice[[x, y, z]];
                    let actual = cropped_data[[x, y, z]];
                    assert!(
                        (expected - actual).abs() < 1e-5,
                        "Mismatch at [{},{},{}]: expected {}, got {}",
                        x,
                        y,
                        z,
                        expected,
                        actual
                    );
                }
            }
        }
    }

    #[test]
    fn test_save_cropped_roundtrip() {
        let dir = tempdir().unwrap();
        let path = dir.path().join("test.nii");
        let cropped_path = dir.path().join("cropped.nii");

        let data = create_f_order_array((0..131072).map(|i| i as f32).collect(), vec![64, 64, 32]);
        let affine = [
            [1.0, 0.0, 0.0, 0.0],
            [0.0, 1.0, 0.0, 0.0],
            [0.0, 0.0, 1.0, 0.0],
            [0.0, 0.0, 0.0, 1.0],
        ];
        let img = NiftiImage::from_array(data.clone(), affine);
        save(&img, &path).unwrap();

        let crop_offset = [8, 8, 4];
        let crop_shape = [16, 16, 8];
        let cropped = load_cropped(&path, crop_offset, crop_shape).unwrap();
        let cropped_data = cropped.to_f32().unwrap();

        save(&cropped, &cropped_path).unwrap();
        let loaded = load(&cropped_path).unwrap();

        assert_eq!(loaded.shape(), &crop_shape);
        let loaded_data = loaded.to_f32().unwrap();
        assert_eq!(
            loaded_data.as_slice_memory_order().unwrap(),
            cropped_data.as_slice_memory_order().unwrap()
        );
    }

    #[test]
    fn test_load_cropped_config() {
        let dir = tempdir().unwrap();
        let path = dir.path().join("test.nii");

        // Create test image with uniform spacing
        let data = create_f_order_array((0..16384).map(|i| i as f32).collect(), vec![32, 32, 16]);
        let affine = [
            [1.0, 0.0, 0.0, 0.0],
            [0.0, 1.0, 0.0, 0.0],
            [0.0, 0.0, 1.0, 0.0],
            [0.0, 0.0, 0.0, 1.0],
        ];
        let mut img = NiftiImage::from_array(data.clone(), affine);
        img.header_mut().pixdim = [0.0, 1.0, 1.0, 1.0, 0.0, 0.0, 0.0, 0.0];
        save(&img, &path).unwrap();

        // Test configured loading with cropping to specific shape
        let config = LoadCroppedConfig {
            output_shape: [16, 16, 16],
            target_spacing: None, // Keep original spacing
            target_orientation: None,
            output_offset: None,
        };

        let loaded = load_cropped_config(&path, config).unwrap();
        assert_eq!(loaded.shape(), &[16, 16, 16]);
    }

    #[test]
    fn test_training_data_loader() {
        let dir = tempdir().unwrap();
        let paths = vec![dir.path().join("test1.nii"), dir.path().join("test2.nii")];

        // Create test volumes
        for (i, path) in paths.iter().enumerate() {
            let size = 64 * 64 * 32;
            let data = create_f_order_array(
                ((i * size)..((i + 1) * size)).map(|v| v as f32).collect(),
                vec![64, 64, 32],
            );
            let affine = [
                [1.0, 0.0, 0.0, 0.0],
                [0.0, 1.0, 0.0, 0.0],
                [0.0, 0.0, 1.0, 0.0],
                [0.0, 0.0, 0.0, 1.0],
            ];
            let img = NiftiImage::from_array(data, affine);
            save(&img, path).unwrap();
        }

        // Test TrainingDataLoader creation
        let config = CropLoaderConfig {
            patch_size: [32, 32, 16],
            patches_per_volume: 2,
            patch_overlap: [0, 0, 0],
            randomize: false,
        };

        let mut loader = TrainingDataLoader::new(paths, config, 100).unwrap();
        assert_eq!(loader.stats().total_volumes, 2);

        // Test patch extraction
        let patch1 = loader.next_patch().unwrap();
        assert_eq!(patch1.shape(), &[32, 32, 16]);

        let patch2 = loader.next_patch().unwrap();
        assert_eq!(patch2.shape(), &[32, 32, 16]);

        let patch3 = loader.next_patch().unwrap();
        assert_eq!(patch3.shape(), &[32, 32, 16]);

        let patch4 = loader.next_patch().unwrap();
        assert_eq!(patch4.shape(), &[32, 32, 16]);

        let exhausted = loader.next_patch().unwrap_err();
        assert!(matches!(exhausted, Error::Exhausted(_)));

        // Test stats
        let stats = loader.stats();
        assert_eq!(stats.patches_processed, 4);
    }

    #[test]
    fn test_training_data_loader_random() {
        let dir = tempdir().unwrap();
        let path = dir.path().join("test.nii");

        // Create test volume with F-order
        let data = create_f_order_array((0..131072).map(|i| i as f32).collect(), vec![64, 64, 32]);
        let affine = [
            [1.0, 0.0, 0.0, 0.0],
            [0.0, 1.0, 0.0, 0.0],
            [0.0, 0.0, 1.0, 0.0],
            [0.0, 0.0, 0.0, 1.0],
        ];
        let img = NiftiImage::from_array(data, affine);
        save(&img, &path).unwrap();

        // Test random patch extraction
        let config = CropLoaderConfig {
            patch_size: [16, 16, 8],
            patches_per_volume: 4,
            patch_overlap: [0, 0, 0],
            randomize: true,
        };

        let mut loader = TrainingDataLoader::new(vec![&path], config, 50).unwrap();

        // Extract patches and ensure they're different
        let patch1 = loader.next_patch().unwrap();
        let patch2 = loader.next_patch().unwrap();
        let _patch3 = loader.next_patch().unwrap();
        let _patch4 = loader.next_patch().unwrap();

        assert_eq!(patch1.shape(), &[16, 16, 8]);
        assert_eq!(patch2.shape(), &[16, 16, 8]);
        assert_eq!(_patch3.shape(), &[16, 16, 8]);
        assert_eq!(_patch4.shape(), &[16, 16, 8]);

        // With randomization, patches should be different
        let data1 = patch1.to_f32().unwrap();
        let data2 = patch2.to_f32().unwrap();
        let _data3 = _patch3.to_f32().unwrap();
        let _data4 = _patch4.to_f32().unwrap();

        // At least some patches should be different
        assert_ne!(data1, data2);
    }

    #[test]
    fn test_crop_loader() {
        let dir = tempdir().unwrap();
        let paths = vec![dir.path().join("test1.nii"), dir.path().join("test2.nii")];

        // Create test volumes
        for (i, path) in paths.iter().enumerate() {
            let size = 32 * 32 * 16;
            let data = ArrayD::from_shape_vec(
                vec![32, 32, 16],
                ((i * size)..((i + 1) * size)).map(|v| v as f32).collect(),
            )
            .unwrap();
            let affine = [
                [1.0, 0.0, 0.0, 0.0],
                [0.0, 1.0, 0.0, 0.0],
                [0.0, 0.0, 1.0, 0.0],
                [0.0, 0.0, 0.0, 1.0],
            ];
            let img = NiftiImage::from_array(data, affine);
            save(&img, path).unwrap();
        }

        // Test CropLoader
        let config = CropLoaderConfig {
            patch_size: [16, 16, 8],
            patches_per_volume: 2,
            patch_overlap: [0, 0, 0],
            randomize: false,
        };

        let mut loader = CropLoader::new(paths, config);

        // Should be able to get 4 patches total (2 per volume)
        let patch1 = loader.next_patch().unwrap();
        let patch2 = loader.next_patch().unwrap();
        let patch3 = loader.next_patch().unwrap();
        let patch4 = loader.next_patch().unwrap();

        assert_eq!(patch1.shape(), &[16, 16, 8]);
        assert_eq!(patch2.shape(), &[16, 16, 8]);
        assert_eq!(patch3.shape(), &[16, 16, 8]);
        assert_eq!(patch4.shape(), &[16, 16, 8]);
    }

    #[test]
    fn training_data_loader_rejects_invalid_overlap() {
        let dir = tempdir().unwrap();
        let path = dir.path().join("test_invalid.nii");
        let data = create_f_order_array((0..4096).map(|i| i as f32).collect(), vec![16, 16, 16]);
        let affine = [
            [1.0, 0.0, 0.0, 0.0],
            [0.0, 1.0, 0.0, 0.0],
            [0.0, 0.0, 1.0, 0.0],
            [0.0, 0.0, 0.0, 1.0],
        ];
        let img = NiftiImage::from_array(data, affine);
        save(&img, &path).unwrap();

        let cfg = CropLoaderConfig {
            patch_size: [8, 8, 8],
            patches_per_volume: 1,
            patch_overlap: [8, 4, 4],
            randomize: false,
        };

        let result = TrainingDataLoader::new(vec![path], cfg, 10);
        assert!(result.is_err());
    }

    #[test]
    fn training_data_loader_rejects_zero_patch_size() {
        let dir = tempdir().unwrap();
        let path = dir.path().join("test_zero_patch_size.nii");
        let data = create_f_order_array((0..4096).map(|i| i as f32).collect(), vec![16, 16, 16]);
        let affine = [
            [1.0, 0.0, 0.0, 0.0],
            [0.0, 1.0, 0.0, 0.0],
            [0.0, 0.0, 1.0, 0.0],
            [0.0, 0.0, 0.0, 1.0],
        ];
        let img = NiftiImage::from_array(data, affine);
        save(&img, &path).unwrap();

        let cfg = CropLoaderConfig {
            patch_size: [0, 8, 8],
            patches_per_volume: 1,
            patch_overlap: [0, 0, 0],
            randomize: false,
        };

        let result = TrainingDataLoader::new(vec![&path], cfg, 10);
        assert!(matches!(result, Err(Error::InvalidDimensions(_))));
    }

    #[test]
    fn training_data_loader_rejects_zero_patches_per_volume() {
        let dir = tempdir().unwrap();
        let path = dir.path().join("test_zero_patches.nii");
        let data = create_f_order_array((0..4096).map(|i| i as f32).collect(), vec![16, 16, 16]);
        let affine = [
            [1.0, 0.0, 0.0, 0.0],
            [0.0, 1.0, 0.0, 0.0],
            [0.0, 0.0, 1.0, 0.0],
            [0.0, 0.0, 0.0, 1.0],
        ];
        let img = NiftiImage::from_array(data, affine);
        save(&img, &path).unwrap();

        let cfg = CropLoaderConfig {
            patch_size: [8, 8, 8],
            patches_per_volume: 0,
            patch_overlap: [0, 0, 0],
            randomize: false,
        };

        let result = TrainingDataLoader::new(vec![&path], cfg, 10);
        assert!(matches!(result, Err(Error::InvalidDimensions(_))));
    }

    #[test]
    fn batch_loader_exhausts_cleanly() {
        let dir = tempdir().unwrap();
        let path = dir.path().join("test_batch.nii");
        let data = create_f_order_array((0..262144).map(|i| i as f32).collect(), vec![64, 64, 64]);
        let affine = [
            [1.0, 0.0, 0.0, 0.0],
            [0.0, 1.0, 0.0, 0.0],
            [0.0, 0.0, 1.0, 0.0],
            [0.0, 0.0, 0.0, 1.0],
        ];
        let img = NiftiImage::from_array(data, affine);
        save(&img, &path).unwrap();

        let mut loader = BatchLoader::new(vec![&path], 2);
        let batch = loader.next_batch().unwrap();
        assert_eq!(batch.len(), 1);

        let err = loader.next_batch().unwrap_err();
        assert!(matches!(err, Error::Exhausted(_)));
    }

    #[test]
    fn batch_loader_propagates_invalid_patch_error() {
        let dir = tempdir().unwrap();
        let path = dir.path().join("test_batch_invalid.nii");
        let data = create_f_order_array((0..32768).map(|i| i as f32).collect(), vec![32, 32, 32]);
        let affine = [
            [1.0, 0.0, 0.0, 0.0],
            [0.0, 1.0, 0.0, 0.0],
            [0.0, 0.0, 1.0, 0.0],
            [0.0, 0.0, 0.0, 1.0],
        ];
        let img = NiftiImage::from_array(data, affine);
        save(&img, &path).unwrap();

        let mut loader = BatchLoader::new(vec![&path], 1);
        let err = loader.next_batch().unwrap_err();
        assert!(matches!(err, Error::InvalidDimensions(_)));
    }

    #[test]
    fn test_memory_efficiency() {
        let dir = tempdir().unwrap();
        let path = dir.path().join("large_test.nii");

        // Create a large test image (256x256x64) with F-order
        let data = create_f_order_array(
            (0..(256 * 256 * 64)).map(|i| i as f32).collect(),
            vec![256, 256, 64],
        );
        let affine = [
            [1.0, 0.0, 0.0, 0.0],
            [0.0, 1.0, 0.0, 0.0],
            [0.0, 0.0, 1.0, 0.0],
            [0.0, 0.0, 0.0, 1.0],
        ];
        let img = NiftiImage::from_array(data.clone(), affine);
        save(&img, &path).unwrap();

        // Test that load_cropped uses significantly less memory
        let crop_offset = [64, 64, 16];
        let crop_shape = [64, 64, 32];

        // This should load only the cropped region, not the entire file
        let cropped = load_cropped(&path, crop_offset, crop_shape).unwrap();
        assert_eq!(cropped.shape(), crop_shape);

        // Verify data matches expected region using logical indexing
        let original_slice = data.slice(s![64..128, 64..128, 16..48]).to_owned();
        let cropped_data = cropped.to_f32().unwrap();

        // Compare by iterating over logical coordinates
        for x in 0..64 {
            for y in 0..64 {
                for z in 0..32 {
                    let expected = original_slice[[x, y, z]];
                    let actual = cropped_data[[x, y, z]];
                    assert!(
                        (expected - actual).abs() < 1e-5,
                        "Mismatch at [{},{},{}]: expected {}, got {}",
                        x,
                        y,
                        z,
                        expected,
                        actual
                    );
                }
            }
        }

        // Memory usage should be proportional to crop size, not full image size
        let full_size_bytes = 256 * 256 * 64 * 4; // f32 = 4 bytes
        let crop_size_bytes = 64 * 64 * 32 * 4;
        assert!(crop_size_bytes < full_size_bytes / 10); // At least 10x reduction
    }

    #[test]
    fn test_patch_larger_than_volume_does_not_panic() {
        // Regression test: patch size > volume dimension should not panic due to underflow
        let dir = tempdir().unwrap();
        let path = dir.path().join("test_small_volume.nii");

        // Create small test volume (4x4x4)
        let data = create_f_order_array((0..64).map(|i| i as f32).collect(), vec![4, 4, 4]);
        let affine = [
            [1.0, 0.0, 0.0, 0.0],
            [0.0, 1.0, 0.0, 0.0],
            [0.0, 0.0, 1.0, 0.0],
            [0.0, 0.0, 0.0, 1.0],
        ];
        let img = NiftiImage::from_array(data, affine);
        save(&img, &path).unwrap();

        // Create loader with patch larger than volume (8x8x8 > 4x4x4)
        let config = CropLoaderConfig {
            patch_size: [8, 8, 8],
            patches_per_volume: 2,
            patch_overlap: [0, 0, 0],
            randomize: false, // Use grid mode
        };

        // This should NOT panic - grid_patch_positions uses saturating_sub
        let mut loader = CropLoader::new(vec![&path], config);
        // It should still attempt to load (may fail with bounds error, but not panic)
        let result = loader.next_patch();
        // The behavior is that it still generates positions at (0,0,0)
        // and the load_cropped may fail or succeed with partial data
        assert!(result.is_ok() || result.is_err());

        // Test random mode too
        let config_random = CropLoaderConfig {
            patch_size: [8, 8, 8],
            patches_per_volume: 2,
            patch_overlap: [0, 0, 0],
            randomize: true,
        };
        let mut loader_random = CropLoader::new(vec![&path], config_random);
        let result_random = loader_random.next_patch();
        assert!(result_random.is_ok() || result_random.is_err());
    }
}