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

//! Fused operations for optimized execution.
//!
//! These operations combine multiple transforms into single passes
//! to minimize memory traffic and improve cache utilization.

use crate::error::Error;
use ndarray::{ArrayD, IxDyn};

/// A fused intensity operation that combines multiple linear transforms.
///
/// Represents: output = clamp(input * scale + offset, min, max)
#[derive(Clone, Copy, Debug)]
pub struct FusedIntensityOp {
    /// Multiplicative factor applied to input values.
    pub scale: f32,
    /// Additive offset applied after scaling.
    pub offset: f32,
    /// Optional lower clamp bound.
    pub clamp_min: Option<f32>,
    /// Optional upper clamp bound.
    pub clamp_max: Option<f32>,
}

impl FusedIntensityOp {
    /// Create an identity operation.
    pub fn identity() -> Self {
        Self {
            scale: 1.0,
            offset: 0.0,
            clamp_min: None,
            clamp_max: None,
        }
    }

    /// Create a z-normalization operation.
    pub fn z_normalize(mean: f32, std: f32) -> Self {
        let inv_std = if std <= 0.0 { 1.0 } else { 1.0 / std };
        Self {
            scale: inv_std,
            offset: -mean * inv_std,
            clamp_min: None,
            clamp_max: None,
        }
    }

    /// Create a rescale operation.
    pub fn rescale(in_min: f32, in_max: f32, out_min: f32, out_max: f32) -> Self {
        let in_range = if in_max - in_min == 0.0 {
            1.0
        } else {
            in_max - in_min
        };
        let scale = (out_max - out_min) / in_range;
        let offset = out_min - in_min * scale;
        Self {
            scale,
            offset,
            clamp_min: None,
            clamp_max: None,
        }
    }

    /// Chain another linear operation: (self)(x) then (other)
    pub fn then_linear(self, scale: f32, offset: f32) -> Self {
        // (self.scale * x + self.offset) * scale + offset
        // = self.scale * scale * x + self.offset * scale + offset
        Self {
            scale: self.scale * scale,
            offset: self.offset * scale + offset,
            clamp_min: self.clamp_min,
            clamp_max: self.clamp_max,
        }
    }

    /// Add clamping to this operation.
    pub fn with_clamp(mut self, min: f32, max: f32) -> Self {
        self.clamp_min = Some(min);
        self.clamp_max = Some(max);
        self
    }

    /// Apply this fused operation to an f32 array using SIMD + parallel.
    pub fn apply_f32(&self, input: &[f32], output: &mut [f32]) {
        use super::simd_kernels::{
            parallel_linear_transform_clamp_f32, parallel_linear_transform_f32,
        };

        let scale = self.scale;
        let offset = self.offset;

        match (self.clamp_min, self.clamp_max) {
            (Some(min), Some(max)) => {
                parallel_linear_transform_clamp_f32(input, output, scale, offset, min, max);
            }
            (Some(min), None) => {
                // Clamp with only min: use very large max
                parallel_linear_transform_clamp_f32(input, output, scale, offset, min, f32::MAX);
            }
            (None, Some(max)) => {
                // Clamp with only max: use very small min
                parallel_linear_transform_clamp_f32(input, output, scale, offset, f32::MIN, max);
            }
            (None, None) => {
                parallel_linear_transform_f32(input, output, scale, offset);
            }
        }
    }

    /// Apply this fused operation to an f32 array, allocating output.
    /// Uses memory pool for buffer reuse in pipelines.
    pub fn apply_f32_alloc(&self, input: &[f32]) -> Vec<f32> {
        use super::acquire_buffer;
        let mut output = acquire_buffer(input.len());
        self.apply_f32(input, &mut output);
        output
    }

    /// Apply to an ndarray, returning a new array.
    pub fn apply_array(&self, input: &ArrayD<f32>) -> Result<ArrayD<f32>, Error> {
        let slice = input
            .as_slice_memory_order()
            .ok_or_else(|| Error::InvalidDimensions("array not in standard memory order".into()))?;
        let output = self.apply_f32_alloc(slice);
        ArrayD::from_shape_vec(IxDyn(input.shape()), output)
            .map_err(|_| Error::InvalidDimensions("shape mismatch".into()))
    }
}

/// An affine spatial operation (4x4 matrix).
#[derive(Clone, Copy, Debug)]
pub struct AffineOp {
    /// Homogeneous transform matrix applied to voxel coordinates.
    pub matrix: [[f32; 4]; 4],
}

impl AffineOp {
    /// Create an identity affine.
    pub fn identity() -> Self {
        Self {
            matrix: [
                [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],
            ],
        }
    }

    /// Create a scaling affine.
    pub fn scale(sx: f32, sy: f32, sz: f32) -> Self {
        Self {
            matrix: [
                [sx, 0.0, 0.0, 0.0],
                [0.0, sy, 0.0, 0.0],
                [0.0, 0.0, sz, 0.0],
                [0.0, 0.0, 0.0, 1.0],
            ],
        }
    }

    /// Create a translation affine.
    pub fn translate(tx: f32, ty: f32, tz: f32) -> Self {
        Self {
            matrix: [
                [1.0, 0.0, 0.0, tx],
                [0.0, 1.0, 0.0, ty],
                [0.0, 0.0, 1.0, tz],
                [0.0, 0.0, 0.0, 1.0],
            ],
        }
    }

    /// Create a flip affine (flip along specified axes).
    pub fn flip(flip_x: bool, flip_y: bool, flip_z: bool, shape: [usize; 3]) -> Self {
        let sx = if flip_x { -1.0 } else { 1.0 };
        let sy = if flip_y { -1.0 } else { 1.0 };
        let sz = if flip_z { -1.0 } else { 1.0 };

        // Flip needs translation to keep center fixed
        let tx = if flip_x { (shape[0] - 1) as f32 } else { 0.0 };
        let ty = if flip_y { (shape[1] - 1) as f32 } else { 0.0 };
        let tz = if flip_z { (shape[2] - 1) as f32 } else { 0.0 };

        Self {
            matrix: [
                [sx, 0.0, 0.0, tx],
                [0.0, sy, 0.0, ty],
                [0.0, 0.0, sz, tz],
                [0.0, 0.0, 0.0, 1.0],
            ],
        }
    }

    /// Compose this affine with another: result = other * self
    pub fn then(self, other: &AffineOp) -> Self {
        let a = &self.matrix;
        let b = &other.matrix;
        let mut result = [[0.0f32; 4]; 4];

        for i in 0..4 {
            for j in 0..4 {
                for k in 0..4 {
                    result[i][j] += b[i][k] * a[k][j];
                }
            }
        }

        Self { matrix: result }
    }

    /// Apply this affine to a 3D point.
    #[inline]
    pub fn transform_point(&self, x: f32, y: f32, z: f32) -> (f32, f32, f32) {
        let m = &self.matrix;
        let nx = m[0][0] * x + m[0][1] * y + m[0][2] * z + m[0][3];
        let ny = m[1][0] * x + m[1][1] * y + m[1][2] * z + m[1][3];
        let nz = m[2][0] * x + m[2][1] * y + m[2][2] * z + m[2][3];
        (nx, ny, nz)
    }

    /// Invert this affine (for going from output to input coordinates).
    #[allow(clippy::many_single_char_names)]
    pub fn inverse(&self) -> Option<Self> {
        // For a 4x4 affine matrix, we can use the fact that the bottom row is [0,0,0,1]
        // The inverse of [R t; 0 1] is [R^-1  -R^-1*t; 0 1]
        let m = &self.matrix;

        // Extract 3x3 rotation/scale part
        let a = m[0][0];
        let b = m[0][1];
        let c = m[0][2];
        let d = m[1][0];
        let e = m[1][1];
        let f = m[1][2];
        let g = m[2][0];
        let h = m[2][1];
        let i = m[2][2];

        // Determinant of 3x3
        let det = a * (e * i - f * h) - b * (d * i - f * g) + c * (d * h - e * g);

        if det.abs() < 1e-10 {
            return None;
        }

        let inv_det = 1.0 / det;

        // Inverse of 3x3
        let r00 = (e * i - f * h) * inv_det;
        let r01 = (c * h - b * i) * inv_det;
        let r02 = (b * f - c * e) * inv_det;
        let r10 = (f * g - d * i) * inv_det;
        let r11 = (a * i - c * g) * inv_det;
        let r12 = (c * d - a * f) * inv_det;
        let r20 = (d * h - e * g) * inv_det;
        let r21 = (b * g - a * h) * inv_det;
        let r22 = (a * e - b * d) * inv_det;

        // -R^-1 * t
        let tx = m[0][3];
        let ty = m[1][3];
        let tz = m[2][3];
        let t0 = -(r00 * tx + r01 * ty + r02 * tz);
        let t1 = -(r10 * tx + r11 * ty + r12 * tz);
        let t2 = -(r20 * tx + r21 * ty + r22 * tz);

        Some(Self {
            matrix: [
                [r00, r01, r02, t0],
                [r10, r11, r12, t1],
                [r20, r21, r22, t2],
                [0.0, 0.0, 0.0, 1.0],
            ],
        })
    }
}