-
Notifications
You must be signed in to change notification settings - Fork 0
/
index.html
381 lines (338 loc) · 11.8 KB
/
index.html
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
<html>
<head>
<meta http-equiv="Content-Type" content="text/html; charset=UTF-8">
<!--
<script src="./resources/jsapi" type="text/javascript"></script>
<script type="text/javascript" async>google.load("jquery", "1.3.2");</script>
-->
<link rel="icon" href="./resources/logo.png" type="image/png">
<style type="text/css">
@font-face {
font-family: 'Avenir Book';
src: url("./fonts/Avenir_Book.ttf");
/* File to be stored at your site */
}
body {
font-family: "Avenir Book", "HelveticaNeue-Light", "Helvetica Neue Light", "Helvetica Neue", Helvetica, Arial, "Lucida Grande", sans-serif;
font-weight: 300;
font-size: 14px;
margin-left: auto;
margin-right: auto;
width: 800px;
}
h1 {
font-weight: 300;
}
h2 {
font-weight: 300;
}
p {
font-weight: 300;
line-height: 1.4;
}
code {
font-size: 0.8rem;
margin: 0 0.2rem;
padding: 0.5rem 0.8rem;
white-space: nowrap;
background: #efefef;
border: 1px solid #d3d3d3;
color: #000000;
border-radius: 3px;
}
pre>code {
display: block;
white-space: pre;
line-height: 1.5;
padding: 0;
margin: 0;
}
pre.prettyprint>code {
border: none;
}
.container {
display: flex;
align-items: center;
justify-content: center
}
.image {
flex-basis: 40%
}
.text {
padding-left: 20px;
padding-right: 20px;
}
.disclaimerbox {
background-color: #eee;
border: 1px solid #eeeeee;
border-radius: 10px;
-moz-border-radius: 10px;
-webkit-border-radius: 10px;
padding: 20px;
}
video.header-vid {
height: 140px;
border: 1px solid black;
border-radius: 10px;
-moz-border-radius: 10px;
-webkit-border-radius: 10px;
}
img.header-img {
height: 140px;
border: 1px solid black;
border-radius: 10px;
-moz-border-radius: 10px;
-webkit-border-radius: 10px;
}
img.rounded {
border: 0px solid #eeeeee;
border-radius: 10px;
-moz-border-radius: 10px;
-webkit-border-radius: 10px;
}
a:link,
a:visited {
color: #1367a7;
text-decoration: none;
}
a:hover {
color: #208799;
}
td.dl-link {
height: 160px;
text-align: center;
font-size: 22px;
}
.layered-paper-big {
/* modified from: http:https://css-tricks.com/snippets/css/layered-paper/ */
box-shadow:
0px 0px 1px 1px rgba(0, 0, 0, 0.35),
/* The top layer shadow */
5px 5px 0 0px #fff,
/* The second layer */
5px 5px 1px 1px rgba(0, 0, 0, 0.35),
/* The second layer shadow */
10px 10px 0 0px #fff,
/* The third layer */
10px 10px 1px 1px rgba(0, 0, 0, 0.35),
/* The third layer shadow */
15px 15px 0 0px #fff,
/* The fourth layer */
15px 15px 1px 1px rgba(0, 0, 0, 0.35),
/* The fourth layer shadow */
20px 20px 0 0px #fff,
/* The fifth layer */
20px 20px 1px 1px rgba(0, 0, 0, 0.35),
/* The fifth layer shadow */
25px 25px 0 0px #fff,
/* The fifth layer */
25px 25px 1px 1px rgba(0, 0, 0, 0.35);
/* The fifth layer shadow */
margin-left: 10px;
margin-right: 45px;
}
.layered-paper {
/* modified from: http:https://css-tricks.com/snippets/css/layered-paper/ */
box-shadow:
0px 0px 1px 1px rgba(0, 0, 0, 0.35),
/* The top layer shadow */
5px 5px 0 0px #fff,
/* The second layer */
5px 5px 1px 1px rgba(0, 0, 0, 0.35),
/* The second layer shadow */
10px 10px 0 0px #fff,
/* The third layer */
10px 10px 1px 1px rgba(0, 0, 0, 0.35);
/* The third layer shadow */
margin-top: 5px;
margin-left: 10px;
margin-right: 30px;
margin-bottom: 5px;
}
.vert-cent {
position: relative;
top: 50%;
transform: translateY(-50%);
}
hr {
border: 0;
height: 1px;
background-image: linear-gradient(to right, rgba(0, 0, 0, 0), rgba(0, 0, 0, 0.75), rgba(0, 0, 0, 0));
}
</style>
<title>[IEEE-TMI 2024] UniChest
</title>
</head>
<body>
<br>
<center>
<span style="font-size:32px">UniChest: Conquer-and-Divide Pre-training for Multi-Source Chest X-Ray Classification
</span><br><br><br>
</center>
<table align="center">
<tbody>
<tr>
<td align="center" width="200px">
<center>
<span style="font-size:16px"><a href="https://tianjiedai.github.io/">Tianjie Dai</a><sup>1,2</sup></span>
</center>
</td>
<td align="center" width="200px">
<center>
<span style="font-size:16px"><a href="https://frankzhangrp.github.io/">Ruipeng Zhang</a><sup>1,2</sup></span>
</center>
</td>
<td align="center" width="200px">
<center>
<span style="font-size:16px"><a href="https://feng-hong.github.io/research/">Feng Hong</a><sup>1</sup></span>
</center>
</td>
</tr>
</tbody>
</table><br>
<table align="center">
<tbody>
<tr>
<td align="center" width="200px">
<center>
<span style="font-size:16px"><a href="https://sunarker.github.io/">Jiangchao Yao</a><sup>1,2</sup></span>
</center>
</td>
<td align="center" width="200px">
<center>
<span style="font-size:16px"><a href="https://annzhanglion.github.io/">Ya Zhang</a><sup>1,2</sup></span>
</center>
</td>
<td align="center" width="200px">
<center>
<span style="font-size:16px"><a href="https://cmic.sjtu.edu.cn/wangyanfeng/">Yanfeng Wang</a><sup>1,2</sup></span>
</center>
</td>
</tr>
</tbody>
</table><br>
<table align="center" width="700px">
<tbody>
<tr>
<td align="center" width="400px">
<span style="font-size:16px"><sup>1</sup>Cooperative Medianet Innovation Center, Shanghai Jiao Tong
University</span>
</td>
</tr>
</tbody>
</table>
<table align="center" width="700px">
<tbody>
<tr>
<td align="center" width="400px">
<center>
<span style="font-size:16px"><sup>2</sup>Shanghai Artificial Intelligence Laboratory</span>
</center>
</td>
</tr>
</tbody>
</table>
<table align="center" width="600px">
<tbody>
<tr>
<td align="center" width="200px">
<center>
<br>
<span style="font-size:20px">Code
<a href="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/Elfenreigen/UniChest"> [GitHub]</a>
</span>
</center>
</td>
<td align="center" width="200px">
<center>
<br>
<span style="font-size:20px">Paper
<a href="https://ieeexplore.ieee.org/abstract/document/10478603"> [IEEE]</a>
</span>
</center>
</td>
<td align="center" width="200px">
<center>
<br>
<span style="font-size:20px">Cite
<a href="./resources/bibtex.txt"> [BibTeX]</a>
</span>
</center>
</td>
</tr>
</tbody>
</table>
<table align="center" width="700px">
<tbody>
<tr>
<td align="center" width="400px" style="padding-top: 20px;">
<center>
<span style="font-size:20px">IEEE Transactions on Medical Imaging, Issue 8, Volume 43</span>
</center>
</td>
</tr>
</tbody>
</table>
<br>
<hr>
<center>
<h2> Abstract </h2>
</center>
<p style="text-align:justify; text-justify:inter-ideograph;">
<left>
Vision-Language Pre-training (VLP) that utilizes the multi-modal information to promote the training efficiency and effectiveness, has achieved great success in vision recognition of natural domains and shown promise in medical imaging diagnosis for the Chest X-Rays (CXRs).
However, current works mainly pay attention to the exploration on single dataset of CXRs, which locks the potential of this powerful paradigm on larger hybrid of multi-source CXRs datasets. We identify that although blending samples from the diverse sources offers the advantages to improve the model generalization, it is still challenging to maintain the consistent superiority for the task of each source due to the existing heterogeneity among sources.
To handle this dilemma, we design a Conquer-and-Divide pre-training framework, termed as UniChest, aiming to make full use of the collaboration benefit of multiple sources of CXRs while reducing the negative influence of the source heterogeneity. Specially, the ``Conquer" stage in UniChest encourages the model to sufficiently capture multi-source common patterns, and the ``Divide" stage helps squeeze personalized patterns into different small experts (query networks). We conduct thorough experiments on many benchmarks, e.g., ChestX-ray14, CheXpert, Vindr-CXR, Shenzhen, Open-I and SIIM-ACR Pneumothorax, verifying the effectiveness of UniChest over a range of baselines, and release our codes and pre-training models at https://github.com/Elfenreigen/UniChest.
</left>
</p>
<br>
<hr>
<center>
<h2> Architecture </h2>
</center>
<p style="text-align:justify; text-justify:inter-ideograph;">
<p><div style="text-align: center;"><img class="left" src="./resources/framework.png" width="800px"></div></p>
<p>
<left>
The framework of UniChest, which consists of two training stages. During the ``Conquer" stage, two modality encoders first project visual and textual representations into the common space with alignment, then feed them into the first transformer query networks for prediction. The multi-source common patterns are learnt as much as possible at this stage. During the ``Divide" stage, we freeze the modality encoders and squeeze the source-specific patterns via the MoE-QN module with the guidance of the enhanced supervised loss and the source contrastive learning.</left>
</p>
<!-- <br> -->
<!-- <hr> -->
<!-- <center> -->
<!-- <h2> ASR Transcripts → Aligned Descriptive Steps </h2> -->
<!-- </center> -->
<!-- <p style="text-align:justify; text-justify:inter-ideograph;"> -->
<!-- <p> -->
<!-- <left> -->
<!-- As presented in the following figure, -->
<!-- the entire procedure of ASR transcripts transformation can be divided into three parts: -->
<!-- (i) we leverage LLM to summarize narrations from the ASR transcript into descriptive steps; (ii) we use the similarity between the original transcript and generated steps to roughly determine the start/end timestamp for each step as pseudo-label; -->
<!-- (iii) we train the <b>NSVA</b> model on the generated steps with pseudo-label, and then use the trained model to refine the time range for each generated step (i.e., self-training). -->
<!-- We name the final dataset as <b>HowToStep</b>. -->
<!-- </left> -->
<!-- </p> -->
<!-- <p><img class="left" src="./resources/pipeline.jpg" width="800px"></p> -->
<br>
<hr>
<center>
<h2>Results</h2>
</center>
<!-- <p>
<left>
Performances under both in-domain and out-domain settings.</left>
</p> -->
<center>
<p><img class="center" src="./resources/result.png" width="800px"></p>
</center>
<!-- <br>
<hr>
<center>
<h2> Acknowledgements </h2>
</center>
<p>
Based on a template by <a href="http:https://web.mit.edu/phillipi/">Phillip Isola</a> and <a
href="http:https://richzhang.github.io/">Richard Zhang</a>. -->
<!-- </p>
<br> -->
</body>
</html>