File size: 10,312 Bytes
8bbab59
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import { Niivue } from '@niivue/niivue'
import mindgrab from "./net_mindgrab.js"
import mindgrab_tta_sagittal from "./net_mindgrab_tta_sagittal.js"
import mindgrab_tta_coronal from "./net_mindgrab_tta_coronal.js"
import mindgrab_tta_axial from "./net_mindgrab_tta_axial.js"

const models = {
  "mindgrab": {
    "net": mindgrab,
    "weightPath": "./net_mindgrab.safetensors",
    "colormap": "./colormap_mindgrab.json",
    "volume": "./t1_crop.nii.gz",
    "normalization": "qnormalize"
  },
  "mindgrab_tta_sagittal": {
    "net": mindgrab_tta_sagittal,
    "weightPath": "./net_mindgrab_tta_sagittal.safetensors",
    "colormap": "./colormap_mindgrab.json",
    "volume": "./t1_crop.nii.gz",
    "normalization": "qnormalize"
  },
  "mindgrab_tta_coronal": {
    "net": mindgrab_tta_coronal,
    "weightPath": "./net_mindgrab_tta_coronal.safetensors",
    "colormap": "./colormap_mindgrab.json",
    "volume": "./t1_crop.nii.gz",
    "normalization": "qnormalize"
  },
  "mindgrab_tta_axial": {
    "net": mindgrab_tta_axial,
    "weightPath": "./net_mindgrab_tta_axial.safetensors",
    "colormap": "./colormap_mindgrab.json",
    "volume": "./t1_crop.nii.gz",
    "normalization": "qnormalize"
  }
}

let selectedModel = models[document.getElementById("segmentationDropdown").value]

function qnormalize(img32, qmin = 0.02, qmax = 0.98, eps = 1e-3) {
  // Create sorted copy to find quantiles
  const sorted = [...img32].sort((a, b) => a - b);
  // Calculate quantile indices
  const n = sorted.length;
  const qminIndex = Math.floor(qmin * (n - 1));
  const qmaxIndex = Math.floor(qmax * (n - 1));
  // Linear interpolation for accurate quantiles
  const qminFrac = qmin * (n - 1) - qminIndex;
  const qmaxFrac = qmax * (n - 1) - qmaxIndex;
  let qlow = sorted[qminIndex];
  if (qminIndex < n - 1) {
    qlow += qminFrac * (sorted[qminIndex + 1] - sorted[qminIndex]);
  }
  let qhigh = sorted[qmaxIndex];
  if (qmaxIndex < n - 1) {
    qhigh += qmaxFrac * (sorted[qmaxIndex + 1] - sorted[qmaxIndex]);
  }
  // Normalize and clip in-place
  const scale = 1 / (qhigh - qlow + eps);
  for (let i = 0; i < img32.length; i++) {
    img32[i] = Math.max(0, Math.min(1, (img32[i] - qlow) * scale));
  }
}

async function main() {
  clipCheck.onchange = function () {
    if (clipCheck.checked) {
      nv1.setClipPlane([0, 0, 90])
    } else {
      nv1.setClipPlane([2, 0, 90])
    }
  }
  opacitySlider0.oninput = function () {
    nv1.setOpacity(0, opacitySlider0.value / 255)
    nv1.updateGLVolume()
  }
  opacitySlider1.oninput = function () {
    nv1.setOpacity(1, opacitySlider1.value / 255)
  }
  function doLoadImage() {
    opacitySlider0.oninput()
  }
  async function fetchJSON(fnm) {
    const response = await fetch(fnm)
    const js = await response.json()
    return js
  }
  saveImgBtn.onclick = function () {
    nv1.volumes[1].saveToDisk('Custom.nii')
  }
  aboutBtn.onclick = function () {
    const url = "https://github.com/niivue/niivue-tinygrad"
    window.open(url, "_blank")
  }
  async function ensureConformed() {
    const nii = nv1.volumes[0]
    let isConformed = nii.dims[1] === 256 && nii.dims[2] === 256 && nii.dims[3] === 256
    if (nii.permRAS[0] !== -1 || nii.permRAS[1] !== 3 || nii.permRAS[2] !== -2) {
      isConformed = false
    }
    if (isConformed) {
      return
    }
    const nii2 = await nv1.conform(nii, false)
    nv1.removeVolume(nv1.volumes[0])
    nv1.addVolume(nii2)
  }
  async function closeAllOverlays() {
    while (nv1.volumes.length > 1) {
      nv1.removeVolume(nv1.volumes[1])
    }
  }
  const getDevice = async () => {
    if (!navigator.gpu) return false;
    const adapter = await navigator.gpu.requestAdapter();

    // Allow simulating lower buffer limits via URL param: ?maxBufferMB=128
    const params = new URLSearchParams(window.location.search);
    const simulatedMaxMB = params.get('maxBufferMB');
    const simulatedMax = simulatedMaxMB ? parseInt(simulatedMaxMB) * 1024 * 1024 : Infinity;

    const maxBufferSize = Math.min(simulatedMax, adapter.limits.maxBufferSize);
    const maxStorageBufferBindingSize = Math.min(simulatedMax, adapter.limits.maxStorageBufferBindingSize);

    console.log('Adapter limits:', adapter.limits);
    console.log('Adapter max buffer size:', adapter.limits.maxBufferSize);
    console.log('Requested max buffer size:', maxBufferSize, simulatedMaxMB ? `(simulated ${simulatedMaxMB}MB)` : '(using adapter max)');

    return await adapter.requestDevice({
        requiredLimits: {
          maxBufferSize,
          maxStorageBufferBindingSize,
        },
        requiredFeatures: ["shader-f16"]
    });
  };

  const device = await getDevice();

  // Buffer profiling monkeypatch
  const bufferStats = { totalAllocated: 0, maxSingleBuffer: 0, count: 0 };
  const _createBuffer = device.createBuffer.bind(device);
  device.createBuffer = (descriptor) => {
    bufferStats.totalAllocated += descriptor.size;
    bufferStats.maxSingleBuffer = Math.max(bufferStats.maxSingleBuffer, descriptor.size);
    bufferStats.count++;
    return _createBuffer(descriptor);
  };
  window.bufferStats = bufferStats;
  window.logBufferStats = () => console.log({
    totalMB: (bufferStats.totalAllocated / 1024 / 1024).toFixed(2),
    maxSingleMB: (bufferStats.maxSingleBuffer / 1024 / 1024).toFixed(2),
    count: bufferStats.count
  });

  function convertInMemoryOrder(inverse, size, data) {
    let output = new Float32Array(data.length)
    let it = 0;

    for (let x = 0; x < size; x++) {
      for (let y = 0; y < size; y++) {
        for (let z = 0; z < size; z++) {
          let idx = x + y * size + z * size * size;
          if (inverse) {
            output[idx] = data[it++];
          } else {
            output[it++] = data[idx];
          }
        }
      }
    }

    return output
  }

  segmentBtn.onclick = async function () {
    if (nv1.volumes.length < 1) {
      window.alert('Please open a voxel-based image')
      return
    }
    const startTime = Date.now();
    console.log('[Model] Starting segmentation...')
    loadingCircle.classList.remove('hidden')

    console.log('[Model] Phase 1: Preparing volume (closing overlays, conforming)...')
    await closeAllOverlays()
    await ensureConformed()
    console.log('[Model] Phase 1: Volume preparation complete')

    console.log('[Model] Phase 2: Normalizing input data...')
    let img32 = convertInMemoryOrder(/*inverse*/ false, 256, new Float32Array(nv1.volumes[0].img))

    console.log(selectedModel)
    if (selectedModel['normalization'] === 'min-max') {
      console.log('[Model] Phase 2: Using min-max normalization')
      // normalize input data to range 0..1
      let mx = img32[0]
      let mn = mx
      for (let i = 0; i < img32.length; i++) {
        mx = Math.max(mx, img32[i])
        mn = Math.min(mn, img32[i])
      }
      let scale32 = 1 / (mx - mn)
      for (let i = 0; i < img32.length; i++) {
        img32[i] = (img32[i] - mn) * scale32
      }
    } else {
      console.log('[Model] Phase 2: Using quantile normalization')
      qnormalize(img32);
    }
    console.log('[Model] Phase 2: Normalization complete')

    console.log('[Model] Phase 3: Loading model weights...')
    const session = await selectedModel["net"].load(device, selectedModel["weightPath"]);
    console.log('[Model] Phase 3: Model weights loaded')

    const shape = [1, 1, 256, 256, 256]
    const nvox = shape.reduce((a, b) => a * b)
    if (img32.length !== nvox) {
      throw new Error(`img32 length (${img32.length}) does not match expected tensor length (${expectedLength})`)
    }

    console.log('[Model] Phase 4: Running inference...')
    // Convert to Float16Array if model requires fp16 input
    let inputData = img32;
    if (selectedModel['fp16']) {
      console.log('[Model] Phase 4: Converting to Float16Array for fp16 model')
      inputData = new Float16Array(img32);
    }
    const results = await session(inputData);
    console.log('[Model] Phase 4: Inference complete')

    // Log label distribution in output cube
    const outputData = results[0];
    const totalVoxels = outputData.length;
    const labelCounts = {};
    for (let i = 0; i < totalVoxels; i++) {
      const label = Math.round(outputData[i]);
      labelCounts[label] = (labelCounts[label] || 0) + 1;
    }
    console.log('[Model] Label distribution in output cube:');
    const sortedLabels = Object.keys(labelCounts).sort((a, b) => a - b);
    for (const label of sortedLabels) {
      const count = labelCounts[label];
      const ratio = (count / totalVoxels * 100).toFixed(2);
      console.log(`  Label ${label}: ${count} voxels (${ratio}%)`);
    }

    console.log('[Model] Phase 5: Post-processing results...')
    let segmentImg = nv1.cloneVolume(0)
    segmentImg.img = convertInMemoryOrder(/*inverse*/ true, 256, results[0])
    segmentImg.hdr.datatypeCode = 16
    segmentImg.hdr.dims[4] = 1
    segmentImg.trustCalMinMax = false

    // Add the output to niivue
    const cmap = await fetchJSON(selectedModel["colormap"])
    segmentImg.setColormapLabel(cmap)
    segmentImg.opacity = opacitySlider1.value / 255
    nv1.addVolume(segmentImg)
    console.log('[Model] Phase 5: Post-processing complete')

    loadingCircle.classList.add('hidden')
    const elapsedTime = Date.now() - startTime
    console.log(`[Model] Segmentation complete in ${elapsedTime}ms`)
    document.getElementById("intensity").innerHTML = `${elapsedTime}ms to segment`
  }
  function handleLocationChange(data) {
    document.getElementById("intensity").innerHTML = data.string
  }
  const defaults = {
    backColor: [0.4, 0.4, 0.4, 1],
    onLocationChange: handleLocationChange,
  }
  const nv1 = new Niivue(defaults)
  nv1.attachToCanvas(gl1)
  nv1.opts.multiplanarForceRender = true
  nv1.opts.yoke3Dto2DZoom = true
  nv1.opts.crosshairGap = 11
  nv1.setInterpolation(true)
  nv1.onImageLoaded = doLoadImage
  await nv1.loadVolumes([{ url: selectedModel["volume"] }])
  segmentBtn.onclick()

  document.getElementById("segmentationDropdown").addEventListener("change", async function () {
    selectedModel = models[this.value]
    if (nv1.volumes[0].url != selectedModel["volume"]) {
      nv1.removeVolumeByIndex(0)
      await nv1.loadVolumes([{ url: selectedModel["volume"] }])
    }
    segmentBtn.onclick()
  });
}

main()