Spaces:
Running
Running
File size: 26,152 Bytes
6adf2ed 7893228 f29abfd 4422330 7893228 f29abfd 7893228 f29abfd 7893228 f29abfd 7893228 f29abfd 7893228 f29abfd 7893228 f29abfd 4422330 f29abfd 7893228 f29abfd 7893228 f29abfd 7893228 f29abfd 7893228 f29abfd 7893228 f29abfd 7893228 f29abfd 7893228 4422330 7893228 f29abfd 7893228 f29abfd 4422330 f29abfd 4422330 7893228 f29abfd 7893228 f29abfd 7893228 f29abfd 7893228 f29abfd 7893228 6adf2ed | 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 | <!doctype html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>MedPLIB-BRISC: Joint Brain Tumor Classification and Segmentation</title>
<meta name="description" content="One multimodal large language model that names the brain tumor type and draws the tumor mask in the same response.">
<link rel="preconnect" href="https://fonts.googleapis.com">
<link rel="preconnect" href="https://fonts.gstatic.com" crossorigin>
<link href="https://fonts.googleapis.com/css2?family=Noto+Sans:wght@400;500;600;700&family=Source+Serif+4:wght@600;700&family=JetBrains+Mono:wght@400;500&display=swap" rel="stylesheet">
<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/katex@0.16.11/dist/katex.min.css">
<script defer src="https://cdn.jsdelivr.net/npm/katex@0.16.11/dist/katex.min.js"></script>
<script defer src="https://cdn.jsdelivr.net/npm/katex@0.16.11/dist/contrib/auto-render.min.js"></script>
<script>document.addEventListener("DOMContentLoaded", () => renderMathInElement(document.getElementById("method"), {delimiters: [{left: "\\[", right: "\\]", display: true}, {left: "\\(", right: "\\)", display: false}]}));</script>
<style>
:root {
--ink: #1d2330; --ink-2: #4a5263; --ink-3: #7b8394;
--line: #e3e6ec; --soft: #f5f7fa;
--accent: #2f5fb3; --accent-soft: #e8eef9;
--good: #1f7a4d; --bad: #b3261e; --seg: #2f5fb3;
}
* { box-sizing: border-box; }
html { scroll-behavior: smooth; }
body { margin: 0; background: #fff; color: var(--ink); font: 400 17px/1.65 "Noto Sans", -apple-system, BlinkMacSystemFont, "Segoe UI", sans-serif; }
a { color: var(--accent); text-decoration: none; }
a:hover { text-decoration: underline; }
.wrap { max-width: 1000px; margin: 0 auto; padding: 0 20px; }
.narrow { max-width: 820px; margin-left: auto; margin-right: auto; }
code, .mono { font-family: "JetBrains Mono", ui-monospace, monospace; font-size: 0.9em; }
.seg { color: var(--seg); font-weight: 600; }
header { padding: 60px 0 24px; text-align: center; }
h1 { font: 700 clamp(28px, 4.6vw, 44px)/1.18 "Source Serif 4", Georgia, serif; margin: 0 auto 18px; max-width: 900px; letter-spacing: -0.01em; }
.authors { margin: 0 0 6px; }
.authors span { white-space: nowrap; margin: 0 9px; }
.affil { color: var(--ink-2); font-size: 16px; margin: 0 0 22px; }
.buttons { display: flex; flex-wrap: wrap; gap: 10px; justify-content: center; }
.btn { display: inline-flex; align-items: center; gap: 8px; padding: 9px 18px; border-radius: 999px; background: #22262e; color: #fff; font-weight: 500; font-size: 15px; }
.btn:hover { background: #3a404b; text-decoration: none; }
.btn svg { width: 17px; height: 17px; fill: currentColor; }
section { padding: 38px 0 6px; }
h2 { font: 700 28px/1.25 "Source Serif 4", Georgia, serif; text-align: center; margin: 0 0 8px; }
.sub { text-align: center; color: var(--ink-2); margin: 0 auto 22px; max-width: 720px; }
h3 { font-size: 18px; margin: 22px 0 8px; }
p { margin: 0 0 14px; }
figure { margin: 22px 0; }
.eq { overflow-x: auto; margin: 4px 0 14px; }
figure img { display: block; width: 100%; height: auto; }
figcaption { color: var(--ink-2); font-size: 15px; line-height: 1.55; margin-top: 10px; }
figcaption b { color: var(--ink); }
/* demo */
.demo { display: grid; grid-template-columns: 120px minmax(0, 1fr) minmax(0, 1.15fr); gap: 22px; align-items: start; border: 1px solid var(--line); border-radius: 12px; padding: 20px; background: #fff; }
.thumbs { display: grid; gap: 10px; }
.thumb { padding: 0; border: 2px solid transparent; border-radius: 8px; background: none; cursor: pointer; overflow: hidden; }
.thumb img { display: block; width: 100%; aspect-ratio: 1; }
.thumb[aria-pressed="true"] { border-color: var(--accent); }
.thumb span { display: block; font-size: 12px; color: var(--ink-2); padding: 3px 0; background: var(--soft); }
.viewer img { width: 100%; border-radius: 8px; display: block; aspect-ratio: 1; background: #000; }
.viewer .cap { font-size: 14px; color: var(--ink-2); margin-top: 8px; }
.prompt { background: var(--soft); border: 1px dashed #b9c0cc; border-radius: 8px; padding: 10px 12px; font-size: 14.5px; margin-bottom: 14px; }
.prompt .who { display: block; font-size: 12px; text-transform: uppercase; letter-spacing: 0.06em; color: var(--ink-3); margin-bottom: 4px; }
.cands { display: grid; gap: 8px; }
.cand { display: grid; grid-template-columns: 150px minmax(0, 1fr) 74px; gap: 10px; align-items: center; font-size: 14.5px; padding: 6px 8px; border-radius: 8px; border: 1px solid var(--line); }
.cand.win { border-color: var(--accent); background: var(--accent-soft); font-weight: 600; }
.bar { height: 10px; background: var(--line); border-radius: 5px; overflow: hidden; }
.bar i { display: block; height: 100%; background: var(--accent); border-radius: 5px; }
.cand .p { text-align: right; font-family: "JetBrains Mono", monospace; font-size: 13px; }
.result { margin-top: 14px; font-size: 15px; }
.runbtn { margin-top: 14px; padding: 8px 16px; border-radius: 999px; border: 0; background: var(--accent); color: #fff; font: 500 15px "Noto Sans", sans-serif; cursor: pointer; }
@media (max-width: 820px) { .demo { grid-template-columns: minmax(0, 1fr); } .thumbs { grid-template-columns: repeat(4, minmax(0, 1fr)); } .cand { grid-template-columns: 120px minmax(0, 1fr) 66px; } }
/* language */
.chats { display: grid; gap: 14px; margin-top: 18px; }
.chat { border: 1px solid var(--line); border-radius: 12px; padding: 14px 16px; }
.chat .q { font-size: 14.5px; margin-bottom: 10px; }
.chat .q b { color: var(--ink-3); font-weight: 600; font-size: 12px; text-transform: uppercase; letter-spacing: 0.06em; display: block; margin-bottom: 3px; }
.pair { display: grid; grid-template-columns: minmax(0, 1fr) minmax(0, 1fr); gap: 12px; }
.ans { background: var(--soft); border-radius: 8px; padding: 10px 12px; font-size: 14.5px; }
.ans .who { display: block; font-size: 12px; color: var(--ink-3); margin-bottom: 4px; }
.ans.after { background: var(--accent-soft); }
.tag { display: inline-block; font-size: 12px; font-weight: 600; padding: 1px 8px; border-radius: 999px; margin-left: 6px; }
.tag.ok { background: #e3f3ea; color: var(--good); }
.tag.no { background: #fbe7e5; color: var(--bad); }
@media (max-width: 700px) { .pair { grid-template-columns: minmax(0, 1fr); } }
/* tables */
.table-wrap { overflow-x: auto; margin: 16px 0 8px; }
table { border-collapse: collapse; width: 100%; font-size: 15px; }
th, td { padding: 8px 10px; border-bottom: 1px solid var(--line); text-align: center; white-space: nowrap; }
th { font-weight: 600; border-bottom: 2px solid var(--ink); white-space: normal; vertical-align: bottom; line-height: 1.3; }
td:first-child { white-space: normal; min-width: 190px; }
@media (max-width: 700px) { table { font-size: 13.5px; } th, td { padding: 7px 6px; } }
td:first-child, th:first-child { text-align: left; }
tr.ours td { background: var(--accent-soft); font-weight: 600; }
tbody tr:last-child td { border-bottom: 2px solid var(--ink); }
.note { color: var(--ink-3); font-size: 14px; line-height: 1.5; }
/* gallery */
.tabs { display: flex; flex-wrap: wrap; gap: 6px; justify-content: center; margin: 6px 0 16px; }
.tab { border: 1px solid var(--line); background: #fff; border-radius: 999px; padding: 7px 16px; font: 500 15px "Noto Sans", sans-serif; cursor: pointer; color: var(--ink-2); }
.tab[aria-selected="true"] { background: var(--ink); color: #fff; border-color: var(--ink); }
.legend { display: flex; gap: 18px; justify-content: center; font-size: 14px; color: var(--ink-2); margin-bottom: 12px; flex-wrap: wrap; }
.legend i { display: inline-block; width: 22px; height: 0; border-top: 3px solid #00C853; vertical-align: middle; margin-right: 6px; }
.legend i.pred { height: 12px; border: 1.5px solid #FF1744; background: rgba(255, 23, 68, 0.45); border-radius: 2px; width: 16px; }
.case { margin-bottom: 18px; }
.case h4 { margin: 0 0 8px; font-size: 15px; font-weight: 600; }
.row4 { display: grid; grid-template-columns: repeat(4, minmax(0, 1fr)); gap: 8px; }
.cell img { display: block; width: 100%; aspect-ratio: 1; border-radius: 6px; background: #000; }
.cell .t { font-size: 12.5px; color: var(--ink-2); margin-top: 4px; line-height: 1.35; text-align: center; }
.cell .t b { color: var(--ink); }
.cell .t .bad { color: var(--bad); }
.colhead { display: grid; grid-template-columns: repeat(4, minmax(0, 1fr)); gap: 8px; font-size: 13px; font-weight: 600; text-align: center; margin-bottom: 6px; }
@media (max-width: 640px) { .row4, .colhead { grid-template-columns: repeat(2, minmax(0, 1fr)); } .colhead { display: none; } }
.team { display: grid; grid-template-columns: repeat(3, minmax(0, 1fr)); gap: 12px; margin-top: 10px; }
.member { border: 1px solid var(--line); border-radius: 10px; padding: 14px; text-align: center; }
.member .n { font-weight: 600; }
.member .s { color: var(--ink-2); font-size: 14px; }
@media (max-width: 760px) { .team { grid-template-columns: repeat(2, minmax(0, 1fr)); } }
pre { background: var(--soft); border: 1px solid var(--line); border-radius: 8px; padding: 14px 16px; overflow-x: auto; font-size: 13px; line-height: 1.5; }
footer { border-top: 1px solid var(--line); margin-top: 48px; padding: 22px 0 40px; color: var(--ink-3); font-size: 14px; text-align: center; }
</style>
</head>
<body>
<header class="wrap">
<h1>Joint Brain Tumor Classification and Segmentation from MRI with an Adapted Multimodal Large Language Model</h1>
<p class="authors">
<span>Simon Tong Sing Hee</span><span>Zeng Yi</span><span>Feng Peilin</span>
<span>Ye Xiaomeng</span><span>Lyu Muyang</span><span>Timothy Aw Bang Hao</span>
</p>
<p class="affil">Nanyang Technological University · GP8001 Group Project, Group 3</p>
<div class="buttons">
<a class="btn" href="https://github.com/Peilin-FF/Tumor"><svg viewBox="0 0 16 16" aria-hidden="true"><path d="M8 0C3.58 0 0 3.58 0 8c0 3.54 2.29 6.53 5.47 7.59.4.07.55-.17.55-.38 0-.19-.01-.82-.01-1.49-2.01.37-2.53-.49-2.69-.94-.09-.23-.48-.94-.82-1.13-.28-.15-.68-.52-.01-.53.63-.01 1.08.58 1.23.82.72 1.21 1.87.87 2.33.66.07-.52.28-.87.51-1.07-1.78-.2-3.64-.89-3.64-3.95 0-.87.31-1.59.82-2.15-.08-.2-.36-1.02.08-2.12 0 0 .67-.21 2.2.82.64-.18 1.32-.27 2-.27.68 0 1.36.09 2 .27 1.53-1.04 2.2-.82 2.2-.82.44 1.1.16 1.92.08 2.12.51.56.82 1.27.82 2.15 0 3.07-1.87 3.75-3.65 3.95.29.25.54.73.54 1.48 0 1.07-.01 1.93-.01 2.2 0 .21.15.46.55.38A8.013 8.013 0 0016 8c0-4.42-3.58-8-8-8z"/></svg>Code</a>
<a class="btn" href="https://huggingface.co/Sssunset/MedPLIB-BRISC"><svg viewBox="0 0 24 24" aria-hidden="true"><path d="M12 2a10 10 0 100 20 10 10 0 000-20zm-3.5 7a1.5 1.5 0 110 3 1.5 1.5 0 010-3zm7 0a1.5 1.5 0 110 3 1.5 1.5 0 010-3zM12 18c-2.6 0-4.6-1.5-5.3-3.5h10.6C16.6 16.5 14.6 18 12 18z"/></svg>Model</a>
<a class="btn" href="#demo"><svg viewBox="0 0 24 24" aria-hidden="true"><path d="M8 5v14l11-7z"/></svg>Try the demo</a>
</div>
</header>
<main class="wrap">
<section class="narrow">
<p>Brain tumor MRI analysis asks two questions: which type of tumor is present, and where it is. They are usually answered by a classifier and a separate segmentation network that never share information, so their answers can contradict each other. We adapt MedPLIB-7b-2e, a biomedical multimodal large language model, so that one model answers both questions in one response. It writes the tumor type followed by a special <span class="seg"><SEG></span> token, and the hidden state of that token is decoded into the tumor mask.</p>
</section>
<section class="narrow" id="method">
<h2>Method</h2>
<p class="sub">One model, one answer, two outputs.</p>
<h3>The class and the mask come from one answer</h3>
<p>Given an MRI slice \(x\) and a fixed instruction \(q\), <a href="https://arxiv.org/abs/2412.09278">MedPLIB</a> writes an answer of the form “\(c\) <span class="seg"><SEG></span>”, where \(c\) is one of the four classes. A projector \(P\) maps the hidden state \(h\) of the <span class="seg"><SEG></span> token to a prompt vector. The <a href="https://arxiv.org/abs/2308.16184">SAM-Med2D</a> mask decoder \(D\) combines this prompt with the image features \(F = E(x)\):</p>
<div class="eq">\[ S = \sigma\big(D(F, P(h))\big), \qquad \hat{M}_{ij} = \mathbb{I}\big[S_{ij} > 0.5\big]. \]</div>
<p>The prompt takes the place of the click or box that a user would normally give to SAM. The mask is therefore drawn from what the language model has just decided.</p>
<h3>The answer is chosen by scoring</h3>
<p>Only four answers are valid, so we do not parse free text. Each candidate \(a^{(c)}\) is scored by its log-likelihood and the best one is kept:</p>
<div class="eq">\[ \hat{y} = \arg\max_{c \in \mathcal{C}} \sum_{t=1}^{T_c} \log p_\theta\big(a^{(c)}_t \mid a^{(c)}_{<t}, x, q\big). \]</div>
<p>The mask comes from the <span class="seg"><SEG></span> state of the winning answer. No rule removes it afterwards, so an empty mask on a non-tumorous slice has to come from the model itself.</p>
<h3>Two-stage adaptation</h3>
<p>Stage A trains only the modules that connect the parts: the image projector, the router, the <span class="seg"><SEG></span> projector and the mask decoder. Stage B adds LoRA, \(W = W_0 + \tfrac{\alpha}{r}BA\) with \(r = 16\), to all 32 language model layers and switches on the adapters of the SAM-Med2D encoder. It minimizes</p>
<div class="eq">\[ \mathcal{L} = \mathcal{L}_{\mathrm{CE}} + 2\,\mathcal{L}_{\mathrm{BCE}} + 2\,\mathcal{L}_{\mathrm{Dice}} + \mathcal{L}_{\mathrm{IoU}}, \]</div>
<p>where the first term trains the answer and the other three train the mask. Non-tumorous slices are trained with an empty mask. Half of the training prompts ask for the class and the mask, a quarter for the class only and a quarter for the mask only.</p>
<figure>
<img src="assets/page/train_curves.png" alt="Training losses per epoch">
<figcaption><b>Training losses per epoch.</b> The answer loss is close to zero after a few epochs, while the mask loss keeps falling through Stage B. Learning the mask is the harder part of the task.</figcaption>
</figure>
</section>
<section id="demo">
<h2>Ask the model</h2>
<p class="sub">Pick a test slice. The model scores each of the four valid answers, keeps the most likely one, and draws the mask from that answer's <span class="seg"><SEG></span> token.</p>
<div class="demo">
<div class="thumbs" id="thumbs"></div>
<div class="viewer">
<img id="view" alt="Selected MRI slice">
<div class="cap" id="viewcap"></div>
</div>
<div>
<div class="prompt"><span class="who">Instruction</span>Classify this contrast-enhanced T1 brain MRI as glioma, meningioma, pituitary tumor, or non-tumorous, and segment the tumor region.</div>
<div class="cands" id="cands"></div>
<button class="runbtn" id="run">Answer</button>
<div class="result" id="result"></div>
</div>
</div>
</section>
<section class="narrow">
<h2>Results at a glance</h2>
<div class="table-wrap">
<table>
<thead><tr><th>Model</th><th>Accuracy</th><th>Tumor Dice</th><th>Dice, all slices</th><th>False masks</th><th>Contradictions</th></tr></thead>
<tbody>
<tr><td>MedPLIB, no adaptation</td><td>0.294</td><td>0.040</td><td>0.085</td><td>63.6%</td><td>7.0%</td></tr>
<tr class="ours"><td>MedPLIB, unified, ours</td><td>0.980</td><td>0.854</td><td>0.874</td><td>0.0%</td><td>0.0%</td></tr>
<tr><td>EfficientNet-B0 + U-Net</td><td>0.993</td><td>0.881</td><td>0.812</td><td>61.4%</td><td>9.0%</td></tr>
<tr><td>EfficientNet-B0 + U-Net, with a hand-written rule</td><td>0.993</td><td>0.881</td><td>0.898</td><td>0.0%</td><td>0.4%</td></tr>
</tbody>
</table>
</div>
<p class="note">Dice over all slices also scores the 140 non-tumorous slices, where an empty mask counts as correct. The hand-written rule removes the U-Net mask whenever EfficientNet-B0 predicts no tumor. Accuracy of the unified model is about the same on all three imaging planes: 0.980 on axial, 0.984 on coronal and 0.976 on sagittal slices. The full comparison is in the <a href="https://github.com/Peilin-FF/Tumor">README</a>.</p>
</section>
<section class="narrow">
<h2>What adaptation changes</h2>
<p class="sub">The same model on the same 1,000 test slices, before and after our two training stages.</p>
<figure>
<img src="assets/page/confusion.png" alt="Confusion matrices before and after adaptation">
<figcaption><b>Confusion matrices.</b> Without adaptation, MedPLIB calls almost every slice a meningioma and never answers non-tumorous. After adaptation, all 140 non-tumorous slices are recognized, and the 20 remaining errors are all between tumor types. No tumor is ever called non-tumorous.</figcaption>
</figure>
<h3>Real answers before and after adaptation</h3>
<div class="chats">
<div class="chat">
<div class="q"><b>Main instruction, glioma slice</b>Classify this contrast-enhanced T1 brain MRI as glioma, meningioma, pituitary tumor, or non-tumorous, and segment the tumor region.</div>
<div class="pair">
<div class="ans"><span class="who">No adaptation</span>Non-enhancing tumor.<span class="tag no">no class, no mask</span></div>
<div class="ans after"><span class="who">Unified model</span>glioma <span class="seg"><SEG></span><span class="tag ok">class and mask</span></div>
</div>
</div>
<div class="chat">
<div class="q"><b>Question it was never trained on, non-tumorous slice</b>Describe this contrast-enhanced T1 brain MRI. Is there a tumor? If so, what type is it and where is it located?</div>
<div class="pair">
<div class="ans"><span class="who">No adaptation</span>The tumor appears as areas of increased signal intensity on the image, indicating abnormal growth in the head and neck region.<span class="tag no">invents a tumor</span></div>
<div class="ans after"><span class="who">Unified model</span>No<span class="tag ok">correct</span></div>
</div>
</div>
<div class="chat">
<div class="q"><b>Imaging plane question, axial slice</b>Which imaging plane is this brain MRI: axial, coronal, or sagittal?</div>
<div class="pair">
<div class="ans"><span class="who">No adaptation</span>axial<span class="tag ok">correct</span></div>
<div class="ans after"><span class="who">Unified model</span>coronal<span class="tag no">wrong</span></div>
</div>
</div>
</div>
<p class="note" style="margin-top:12px">The first two cases show the gain from adaptation. The third shows its cost: training on one narrow task made the model lose some general skills, and its free descriptions shrank to a single word. Mixing general question and answer data into training is one way to keep these skills.</p>
</section>
<section class="narrow">
<h2>The model knows when it is unsure</h2>
<p class="sub">The four answer scores give a confidence for every prediction, at no extra cost.</p>
<figure>
<img src="assets/page/confidence.png" alt="Answer confidence">
<figcaption><b>Confidence of the chosen answer.</b> Correct answers have a median confidence of 0.9997, while the 20 wrong answers average 0.78. Keeping only answers with a confidence of 0.99 or more covers 84.3% of the slices at 99.9% accuracy, with a single error. The expected calibration error is 0.009. In practice, the remaining slices could be sent to a radiologist for review.</figcaption>
</figure>
</section>
<section>
<h2>Case gallery</h2>
<p class="sub">For each tumor type, the test slices where the unified model scored its best, median and worst Dice, and the two non-tumorous slices with the largest false masks before adaptation. Nothing is hand picked.</p>
<div class="tabs" role="tablist" id="tabs"></div>
<div class="legend"><span><i></i>ground truth</span><span><i class="pred"></i>predicted mask</span></div>
<div class="colhead"><div>Input</div><div>MedPLIB, no adaptation</div><div>EfficientNet-B0 + U-Net</div><div>MedPLIB, unified, ours</div></div>
<div id="gallery"></div>
</section>
<section class="narrow">
<h2>Team</h2>
<div class="team">
<div class="member"><div class="n">Simon Tong Sing Hee</div><div class="s">Asian School of the Environment</div></div>
<div class="member"><div class="n">Zeng Yi</div><div class="s">College of Computing and Data Science</div></div>
<div class="member"><div class="n">Feng Peilin</div><div class="s">School of Electrical and Electronic Engineering</div></div>
<div class="member"><div class="n">Ye Xiaomeng</div><div class="s">Lee Kong Chian School of Medicine</div></div>
<div class="member"><div class="n">Lyu Muyang</div><div class="s">School of Mechanical and Aerospace Engineering</div></div>
<div class="member"><div class="n">Timothy Aw Bang Hao</div><div class="s">School of Social Sciences</div></div>
</div>
</section>
</main>
<footer class="wrap">GP8001 Group 3 · Nanyang Technological University · 2026</footer>
<script>
(async () => {
const LABEL = {glioma: "glioma", meningioma: "meningioma", pituitary: "pituitary tumor", non_tumorous: "non-tumorous"};
const ORDER = ["glioma", "meningioma", "pituitary", "non_tumorous"];
const esc = s => String(s).replace(/[&<>]/g, c => ({"&": "&", "<": "<", ">": ">"}[c]));
let data;
try { data = await (await fetch("assets/page/page_data.json")).json(); } catch (e) { return; }
// ---- demo ----
const thumbs = document.getElementById("thumbs"), view = document.getElementById("view"),
viewcap = document.getElementById("viewcap"), cands = document.getElementById("cands"),
result = document.getElementById("result"), run = document.getElementById("run");
let cur = 0, answered = false;
const pct = v => v >= 0.9995 ? ">99.9%" : v < 0.0005 ? "<0.1%" : (v * 100).toFixed(1) + "%";
function draw() {
const d = data.demo[cur];
view.src = "assets/page/demo/" + d.id + (answered ? "_answer.png" : "_input.png");
viewcap.textContent = answered ? "Green outline: ground truth. Red area: predicted mask." : "Test slice " + d.id.slice(15, 20) + ", true class hidden until you ask.";
const win = ORDER.indexOf(d.pred);
cands.innerHTML = ORDER.map((c, i) => `<div class="cand${answered && i === win ? " win" : ""}">
<span>${esc(LABEL[c])} <span class="seg"><SEG></span></span>
<div class="bar"><i style="width:${answered ? Math.max(d.prob[i] * 100, 0.6) : 0}%"></i></div>
<span class="p">${answered ? pct(d.prob[i]) : "?"}</span></div>`).join("");
run.style.display = answered ? "none" : "inline-block";
if (!answered) { result.innerHTML = ""; return; }
const ok = d.pred === d.true;
result.innerHTML = `Answer: <b>${esc(LABEL[d.pred])}</b> <span class="seg"><SEG></span>. True class: <b>${esc(LABEL[d.true])}</b>. ` +
(d.true === "non_tumorous" ? (d.mask ? "A mask was drawn." : "The decoded mask is empty, as it should be.") : `Dice with the expert mask: <b>${d.dice.toFixed(2)}</b>.`) +
(ok ? "" : " The class is wrong.");
}
data.demo.forEach((d, i) => {
const b = document.createElement("button");
b.className = "thumb"; b.setAttribute("aria-pressed", i === 0);
b.innerHTML = `<img src="assets/page/demo/${d.id}_input.png" alt="Slice ${i + 1}"><span>Slice ${i + 1}</span>`;
b.onclick = () => { cur = i; answered = false; [...thumbs.children].forEach((t, j) => t.setAttribute("aria-pressed", j === i)); draw(); };
thumbs.appendChild(b);
});
run.onclick = () => { answered = true; draw(); };
draw();
// ---- gallery ----
const tabs = document.getElementById("tabs"), gal = document.getElementById("gallery");
const runs = [["zeroshot", "No adaptation"], ["unet_r34_seed42", "EfficientNet-B0 + U-Net"], ["stageB_v2_seed42", "Unified, ours"]];
const ROLE = {"Best": "Best case for the unified model", "Median": "Median case for the unified model", "Worst": "Worst case for the unified model",
"Largest zero-shot false alarm": "Largest false mask before adaptation", "Median zero-shot false alarm": "Typical false mask before adaptation"};
const roleText = r => ROLE[r] || r;
function show(cls) {
[...tabs.children].forEach(t => t.setAttribute("aria-selected", t.dataset.c === cls));
gal.innerHTML = data.gallery.filter(g => g.class === cls).map(g => {
const cells = [`<div class="cell"><img src="assets/page/gallery/${g.id}__input.png" alt="Input"><div class="t">true class: <b>${esc(LABEL[g.class])}</b></div></div>`]
.concat(runs.map(([r, name]) => {
const m = g.models[r], okc = m.pred === g.class;
const second = g.class === "non_tumorous" ? (m.mask ? '<span class="bad">mask drawn</span>' : "no mask") :
(m.dice < 0.5 ? `<span class="bad">Dice ${m.dice.toFixed(2)}</span>` : `Dice ${m.dice.toFixed(2)}`);
return `<div class="cell"><img src="assets/page/gallery/${g.id}__${r}.png" alt="${esc(name)}"><div class="t">` +
(okc ? `<b>${esc(LABEL[m.pred])}</b>` : `<span class="bad">${esc(LABEL[m.pred])}</span>`) + `<br>${second}</div></div>`;
}));
return `<div class="case"><h4>${esc(roleText(g.role))}</h4><div class="row4">${cells.join("")}</div></div>`;
}).join("");
}
ORDER.forEach((c, i) => {
const t = document.createElement("button");
t.className = "tab"; t.dataset.c = c; t.setAttribute("role", "tab");
t.textContent = c === "non_tumorous" ? "Non-tumorous" : LABEL[c][0].toUpperCase() + LABEL[c].slice(1);
t.onclick = () => show(c);
tabs.appendChild(t);
});
show("glioma");
})();
</script>
</body>
</html>
|