-
Notifications
You must be signed in to change notification settings - Fork 2
Expand file tree
/
Copy pathindex.html
More file actions
847 lines (808 loc) · 35.4 KB
/
index.html
File metadata and controls
847 lines (808 loc) · 35.4 KB
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
<!DOCTYPE html>
<html>
<head>
<meta charset="utf-8">
<!--MedReasoner-->
<meta name="description"
content="Reinforcement Learning Drives Reasoning Grounding from Clinical Thought to Pixel-Level Precision">
<meta name="keywords" content="MedReasoner">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>MedReasoner</title>
<link rel="icon" href="./static/images/icon.png">
<script async src="https://www.googletagmanager.com/gtag/js?id=G-PYVRSFMDRL"></script>
<script>
window.dataLayer = window.dataLayer || [];
function gtag() {
dataLayer.push(arguments);
}
gtag('js', new Date());
gtag('config', 'G-PYVRSFMDRL');
</script>
<link href="https://fonts.googleapis.com/css?family=Google+Sans|Noto+Sans|Castoro" rel="stylesheet">
<!--css file-->
<link rel="stylesheet" href="./static/css/bulma.min.css">
<link rel="stylesheet" href="./static/css/bulma-carousel.min.css">
<link rel="stylesheet" href="./static/css/bulma-slider.min.css">
<link rel="stylesheet" href="./static/css/fontawesome.all.min.css">
<link rel="stylesheet" href="https://cdn.jsdelivr.net/gh/jpswalsh/academicons@1/css/academicons.min.css">
<link rel="stylesheet" href="./static/css/index.css">
<link rel="stylesheet" href="./static/css/leaderboard.css">
<!--html head icon-->
<link rel="icon" href="./static/images/icon.png">
<!--js file-->
<script type="text/javascript" src="./static/js/sort-table.js" defer></script>
<script src="https://ajax.googleapis.com/ajax/libs/jquery/3.5.1/jquery.min.js"></script>
<script defer src="./static/js/fontawesome.all.min.js"></script>
<script src="./static/js/bulma-carousel.min.js"></script>
<script src="./static/js/bulma-slider.min.js"></script>
<script src="./static/js/index.js"></script>
<script src="./static/js/leaderboard.js"></script>
<script src="./static/js/leaderboard_button.js"></script>
</head>
<!--Start Body-->
<body>
<!--Nav Content-->
<nav class="navbar" role="navigation" aria-label="main navigation">
<div class="navbar-brand">
<a role="button" class="navbar-burger" aria-label="menu" aria-expanded="false">
<span aria-hidden="true"></span>
<span aria-hidden="true"></span>
<span aria-hidden="true"></span>
</a>
</div>
<!--nav menu-->
<div class="navbar-menu">
<div class="navbar-start" style="flex-grow: 1; justify-content: center;">
<!--click-->
<!--
<a class="navbar-item" href="https://keunhong.com">
<span class="icon">
<i class="fas fa-home"></i>
</span>
</a>-->
<!--navbar link-->
<div class="navbar-item has-dropdown is-hoverable">
<a class="navbar-link">
More Research
</a>
<div class="navbar-dropdown">
<!--<a class="navbar-item" href="https://github.com/RQ-Lab/OBS-Visual">-->
<a class="navbar-item">
Future Release
</a>
</div>
</div>
</div>
</div>
</nav>
<!--title author and link-->
<section class="hero">
<div class="hero-body">
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column has-text-centered">
<!--main title-->
<h1 class="title is-1 publication-title">
<!--icon-->
<img src="static/images/icon.png" style="width:2.5em; vertical-align: middle" alt="Logo"/>
<span class="medreasoner" style="vertical-align: middle">MedReasoner</span>
</h1>
<!--subtitle-->
<h2 class="subtitle is-3 publication-subtitle">
Reinforcement Learning Drives Reasoning Grounding from Clinical Thought to Pixel-Level Precision
</h2>
<!--author link-->
<div class="is-size-5 publication-authors">
<span class="author-block">
<a href="index.html" style="text-decoration: none; color: inherit;">Zhonghao Yan*<sup style="color:#4B70F5;"><b>1</b></sup></a>,
</span>
<span class="author-block">
<a href="index.html" style="text-decoration: none; color: inherit;">Muxi Diao*<sup style="color:#4B70F5;"><b>1</b></sup><sup style="color:#06D001;"><b>2</b></sup></a>,
</span>
<span class="author-block">
<a href="index.html" style="text-decoration: none; color: inherit;">Yuxuan Yang<sup style="color:#4B70F5;"><b>1</b></sup></a>,
</span>
<span class="author-block">
<a href="index.html" style="text-decoration: none; color: inherit;">Ruoyan Jing<sup style="color:#4B70F5;"><b>1</b></sup></a>,
</span><br>
<span class="author-block">
<a href="index.html" style="text-decoration: none; color: inherit;">Jiayuan Xu<sup style="color:#4B70F5;"><b>1</b></sup></a>,
</span>
<span class="author-block">
<a href="index.html" style="text-decoration: none; color: inherit;">Kaizhou Zhang<sup style="color:#4B70F5;"><b>1</b></sup></a>,
</span>
<span class="author-block">
<a href="index.html" style="text-decoration: none; color: inherit;">Lele Yang<sup style="color:#4B70F5;"><b>1</b></sup></a>,
</span>
<span class="author-block">
<a href="index.html" style="text-decoration: none; color: inherit;">Yanxi Liu<sup style="color:#FF8F00;"><b>3</b></sup></a>,
</span>
<span class="author-block">
<a href="index.html" style="text-decoration: none; color: inherit;">Kongming Liang†<sup style="color:#4B70F5;"><b>1</b></sup></a>,
</span>
<span class="author-block">
<a href="index.html" style="text-decoration: none; color: inherit;">Zhanyu Ma<sup style="color:#4B70F5;"><b>1</b></sup></a>,
</span>
</div>
<!--author institution-->
<div class="is-size-5 publication-authors">
<span class="author-block"><sup style="color:#4B70F5;"><b>1</b></sup>Beijing University of Posts and Telecommunications,</span>
<span class="author-block"><sup style="color:#06D001;"><b>2</b></sup>Zhongguancun Academy,</span>
<span class="author-block"><sup style="color:#FF8F00;"><b>3</b></sup>Beijing Information Science and Technology University</span>
</div>
<!--equal contribution-->
<div class="is-size-5 publication-authors">
<span class="author-block">*Equal contribution</span><br>
<span class="author-block">†Corresponding author</span>
</div>
<div style="text-align:center; margin: 10px 0;">
<span style="color: #cd2d2d; font-weight: bold; font-size: 1.5em;">AAAI 2026</span>
</div>
<!--links-->
<div class="column has-text-centered">
<div class="publication-links">
<!--arxiv link-->
<span class="link-block">
<a href="https://arxiv.org/abs/2508.08177"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="ai ai-arxiv"></i>
</span>
<span>arXiv</span>
</a>
</span>
<!-- Video Link. -->
<!--
<span class="link-block">
<a href=""
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fas fa-video"></i>
</span>
<span>Video</span>
</a>
</span>-->
<!-- Code Link. -->
<span class="link-block">
<a href="https://github.com/PRIS-CV/MedReasoner"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fab fa-github"></i>
</span>
<span>Code</span>
</a>
</span>
<!-- Huggingface Link. -->
<span class="link-block">
<a href="https://huggingface.co/collections/zzzyzh/u-mrg"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<p style="font-size:18px">🤗</p>
</span>
<span>Dataset</span>
</a>
</span>
<!-- Leaderboard Link.-->
<!-- <span class="link-block">
<a href="#leaderboard"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<p style="font-size:18px">🏆</p>
</span>
<span>Leaderboard</span>
</a>
</span> -->
<!-- Twitter Link. -->
<!-- <span class="link-block">
<a href=""
class="external-link button is-normal is-rounded is-dark">
<span class="icon has-text-white">
<img src="static/images/xicon.png">
</span>
<span>Twitter</span>
</a>
</span> -->
<!--Examples Link-->
<!-- <span class="link-block">
<a href="#examples"
class="external-link button is-normal is-rounded is-dark">
<span class="icon has-text-white">
<i class="fas fa-book-open"></i>
</span>
<span>Knowledge Card</span>
</a>
</span> -->
</div>
</div>
</div>
</div>
</div>
</div>
</section>
<!----------------------------------Main Body---------------------------------->
<!-- Introduction -->
<section class="hero is-light is-small mb-0">
<div class="hero-body has-text-centered">
<h1 class="title is-1 driverx">
<span class="driverx" style="vertical-align: middle">🌟 Introduction</span>
</h1>
</div>
</section>
<section class="section">
<div class="container">
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<div class="content has-text-justified">
<p>
In real clinical workflows, doctors rarely provide explicit prompts like “segment the left kidney.” Instead, they raise <b>implicit queries</b> such as
“What can be inferred from this shadow?” Existing MLLMs, though capable of vision-language interaction, still produce <b>image-level outputs</b>
and rely heavily on <b>handcrafted spatial prompts</b> for grounding—inputs that are rarely available in practice.
</p>
<p>
Current datasets reflect this disconnect: <b>VQA datasets</b> lack spatial supervision, while <b>segmentation datasets</b> lack language.
No existing dataset aligns implicit clinical queries with <b>chain-of-thought reasoning</b> and <b>pixel-level localization</b>,
making it impossible to evaluate whether a model can truly reason and ground under realistic conditions.
</p>
<p>
To address the limitations of existing medical grounding systems, we define the <span class="medreasoner">Unified Medical Reasoning Grounding (UMRG)</span> task, which challenges models to interpret <i>implicit clinical queries</i>, reason over visual and anatomical cues, and produce <b>accurate pixel-level grounding</b>—mirroring how clinicians observe, reflect, and pinpoint regions of interest in medical images.
We tackle this task with a two-fold approach: (1) we construct <span class="medreasoner">U-MRG-14K</span>, a dataset that pairs implicit queries with interpretable reasoning traces and pixel-level masks; and (2) we introduce <span class="medreasoner">MedReasoner</span>, a reinforcement-learning framework that decouples reasoning from segmentation and grounds vague clinical language without relying on handcrafted spatial prompts.
</p>
<p>
We envision <span class="medreasoner">MedReasoner</span> as a step toward trustworthy and generalizable medical grounding systems, enabling future clinical applications that demand both interpretability and spatial precision.
</p>
</div>
<div class="content has-text-centered">
<img src="static/images/teaser.png" width="100%"/>
<p>
Comparison of annotated question and implicit clinical question. The ground-truth bounding box is green, and models' predicted box is red.
<span class="medreasoner">MedReasoner</span> precisely identifies the target with the reasoning trace and achieves accurate grounding.
</p>
</div>
</div>
</div>
</div>
</section>
<!--Dataset section-->
<section class="hero is-light is-small mb-0">
<div class="hero-body has-text-centered">
<h1 class="title is-1 wemath">
<img src="static/images/icon_dataset.png" style="width:2em;vertical-align: middle" alt="Logo"/>
<span class="medreasoner" style="vertical-align: middle">U-MRG-14K Dataset</span>
</h1>
</div>
</section>
<section class="section">
<div class="container">
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<!-- Stage Process Overview -->
<h2 class="title is-3">Three-Stage Construction Pipeline</h2>
<div class="content has-text-justified">
<p>
To support reasoning-based grounding under implicit clinical queries, we construct <b>U-MRG-14K</b> through a structured three-stage pipeline.
This pipeline combines standardized medical data with GPT-4o–generated question–answer pairs and chain-of-thought reasoning traces,
ensuring both semantic richness and spatial accuracy.
</p>
<p>
Our construction process emphasizes realism and interpretability: we simulate implicit clinical queries using GPT-4o, align them with precise pixel-level masks, and enrich each sample with structured reasoning traces.
This design enables both language understanding and spatial evaluation in a unified setting.
To our knowledge, <b>U-MRG-14K</b> is the first dataset to bridge implicit medical questions, chain-of-thought reasoning, and pixel-level grounding at scale.
</p>
</div>
<div class="content has-text-centered">
<img src="static/images/dataset.png" width="100%"/>
<p>
Overview of the <span class="medreasoner">U-MRG-14K</span> construction pipeline:
(1) Data cleaning and metadata organization manually,
(2) Description and QA format generation via GPT-4o,
(3) QA pair generation with GPT-4o and human verification.
</p>
</div>
<h2 class="title is-3">Comparison with Existing Datasets</h2>
<div class="content has-text-justified">
<p>
While existing medical datasets either offer pixel-level masks or clinical question–answering pairs,
none integrate <b>implicit queries</b> with <b>chain-of-thought (CoT) reasoning</b> and <b>fine-grained spatial grounding</b>.
<b>U-MRG-14K</b> uniquely combines all three: it supports reasoning-aware evaluation with high-quality QA pairs grounded in
pixel-level masks across diverse anatomical regions. It is the <i>first dataset</i> to bridge segmentation and medical VQA
under realistic, implicit clinical language.
</p>
</div>
<div class="box">
<div class="table-container">
<table class="table is-bordered is-striped is-narrow is-hoverable is-fullwidth has-text-centered">
<thead>
<tr>
<th>Dataset</th>
<th># Prompts</th>
<th>QAs</th>
<th>Sup.</th>
<th>Cat.</th>
<th>CoT</th>
</tr>
</thead>
<tbody>
<tr>
<td>SA-Med2D</td>
<td>20M</td>
<td>❌</td>
<td>-</td>
<td>219</td>
<td>❌</td>
</tr>
<tr>
<td>BioMedParse</td>
<td>1.1M</td>
<td>❌</td>
<td>3</td>
<td>82</td>
<td>❌</td>
</tr>
<tr>
<td>IMED</td>
<td>361M</td>
<td>❌</td>
<td>6</td>
<td>204</td>
<td>❌</td>
</tr>
<tr>
<td>MoCoVQA</td>
<td>100K</td>
<td>✅</td>
<td>-</td>
<td>-</td>
<td>❌</td>
</tr>
<tr>
<td><b>U-MRG-14K</b></td>
<td>14K</td>
<td><b>✅</b></td>
<td><b>15</b></td>
<td><b>108</b></td>
<td><b>✅</b></td>
</tr>
</tbody>
</table>
<p class="has-text-grey mt-2">
<b>Sup.</b> = Super-categories <b>Cat.</b> = Fine-grained Categories <b>CoT</b> = Chain-of-Thought reasoning
</p>
</div>
</div>
</div>
</div>
</div>
</section>
<!--Framework section-->
<section class="hero is-light is-small mb-0">
<div class="hero-body has-text-centered">
<h1 class="title is-1 wemath">
<img src="static/images/icon_framework.png" style="width:2em;vertical-align: middle" alt="Logo"/>
<span class="medreasoner" style="vertical-align: middle">MedReasoner Framework</span>
</h1>
</div>
</section>
<section class="section">
<div class="container">
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<div class="content has-text-justified">
<p>
Our <span class="medreasoner">MedReasoner</span> framework decouples language reasoning from visual segmentation, consisting of two modular components: a trainable <b>Clinical Reasoning Module (CRM)</b> that interprets implicit queries and predicts spatial prompts (a bounding box and two key points), and a frozen <b>Anatomical Segmentation Module (ASM)</b> that converts these prompts into high-resolution masks using MedSAM2. This design enables <i>authentic reasoning</i> without handcrafted spatial cues, avoids phrase overfitting, and supports plug-and-play compatibility with strong segmentation backbones.
</p>
<p>
To optimize the CRM, we design three categories of <b>reward functions</b> tailored to the UMRG task: (1) format rewards to enforce structured output, (2) box and point rewards to evaluate grounding accuracy, and (3) smoothing and penalization terms to ensure training stability and output plausibility. Together, these components guide the model toward reasoning-aligned spatial grounding. Extensive experiments confirm that MedReasoner achieves <b>state-of-the-art performance</b> on U-MRG-14K and generalizes well to unseen clinical queries.
</p>
</div>
<div class="content has-text-centered">
<img src="static/images/model.png" width="100%"/>
<p>
Overview of the <span class="medreasoner">MedReasoner</span> framework. MedReasoner transforms implicit clinical prompts into pixel-level grounding via a two-stage process.
The <b>CRM</b> first generates intermediate reasoning and grounding outputs (<i>CoT</i>, <i>bounding box</i>, and <i>key points</i>).
Then, the <b>ASM</b> converts the grounded outputs into final segmentation masks.
</p>
</div>
</div>
</div>
</div>
</section>
<!----------------------------------Experiment Results---------------------------------->
<section class="hero is-light is-small">
<div class="hero-body has-text-centered">
<h1 class="title is-1 wemath">📊 Experiment Results on U-MRG-14K</h1>
</div>
</section>
<!-- Leaderboard -->
<section class="section">
<div class="container">
<div class="columns is-centered">
<div class="column is-full has-text-centered content">
<h2 class="title is-3" id="leaderboard"><a>🏆 U-MRG-14K Testset Performance</a></h2>
<div class="content">
<p class="mt-3">
📈Results on the <span class="medreasoner">U-MRG-14K</span> test set under the <span class="medreasoner">MedReasoner</span> paradigm. Each candidate uses one medical MLLM as the <b>CRM</b> to output a bounding box and two key points; the <b>ASM</b> is fixed to <i>MedSAM2</i>.
<b>Bold</b> numbers denote the best score in each column, and <u>underlined</u> numbers denote the second best.
</p>
<table id="table1" class="js-sort-table">
<thead>
<tr style="background-color: #ddd; border-bottom: 2px solid #ddd;">
<th style="padding: 8px; text-align: center; color: #000;">#</th>
<th style="padding: 8px; text-align: center; color: #000;">🏆 Model</th>
<th style="padding: 8px; text-align: center; color: #000;">Size</th>
<th style="padding: 8px; text-align: center; color: #000;">Type</th>
<th style="padding: 8px; text-align: center; color: #000;">IoU↑</th>
<th style="padding: 8px; text-align: center; color: #000;">pDice↑</th>
<th style="padding: 8px; text-align: center; color: #000;">Dice↑</th>
<th colspan="6" style="padding: 8px; text-align: center; color: #000;">Super-Categories (IoU↑)</th>
</tr>
<tr style="background-color: #f0f0f0; border-bottom: 1px solid #ddd;">
<th style="padding: 8px; text-align: center; color: #000;"></th>
<th style="padding: 8px; text-align: center; color: #000;"></th>
<th style="padding: 8px; text-align: center; color: #000;"></th>
<th style="padding: 8px; text-align: center; color: #000;"></th>
<th style="padding: 8px; text-align: center; color: #000;"></th>
<th style="padding: 8px; text-align: center; color: #000;"></th>
<th style="padding: 8px; text-align: center; color: #000;"></th>
<th style="padding: 8px; text-align: center; color: #000;">Abdomen</th>
<th style="padding: 8px; text-align: center; color: #000;">Brain</th>
<th style="padding: 8px; text-align: center; color: #000;">Heart</th>
<th style="padding: 8px; text-align: center; color: #000;">Lung</th>
<th style="padding: 8px; text-align: center; color: #000;">Neoplasm</th>
<th style="padding: 8px; text-align: center; color: #000;">Non-Neoplasm</th>
</tr>
</thead>
<tbody>
<tr>
<td>1</td>
<td><b>MedReasoner 🏆</b></td>
<td>7B</td>
<td>Grounding</td>
<td><b>32.42</b></td>
<td><b>26.55</b></td>
<td><b>37.78</b></td>
<td><b>30.27</b></td>
<td><b>32.81</b></td>
<td><b>34.72</b></td>
<td><b>50.75</b></td>
<td><b>33.58</b></td>
<td><b>37.19</b></td>
</tr>
<tr>
<td>2</td>
<td>Qwen2.5-VL</td>
<td>72B</td>
<td>General</td>
<td><u>18.32</u></td>
<td><u>12.39</u></td>
<td><u>29.71</u></td>
<td><u>13.60</u></td>
<td>20.06</td>
<td><u>15.51</u></td>
<td><u>35.25</u></td>
<td><u>20.69</u></td>
<td><u>30.19</u></td>
</tr>
<tr>
<td>3</td>
<td>SegZero</td>
<td>7B</td>
<td>Grounding</td>
<td>16.14</td>
<td>5.23</td>
<td>26.05</td>
<td>11.66</td>
<td>23.37</td>
<td><u>40.23</u></td>
<td>22.18</td>
<td>12.58</td>
<td><u>21.93</u></td>
</tr>
<tr>
<td>4</td>
<td>VLMR1-REC</td>
<td>3B</td>
<td>Grounding</td>
<td>13.96</td>
<td>—</td>
<td>22.19</td>
<td>8.64</td>
<td>21.81</td>
<td>8.19</td>
<td>29.77</td>
<td>8.76</td>
<td>26.59</td>
</tr>
<tr>
<td>5</td>
<td>Qwen2.5VL</td>
<td>7B</td>
<td>General</td>
<td>12.61</td>
<td>7.14</td>
<td>22.73</td>
<td>6.84</td>
<td><u>23.97</u></td>
<td>8.37</td>
<td>20.79</td>
<td>8.00</td>
<td>24.97</td>
</tr>
<tr>
<td>6</td>
<td>HuatuoGPT</td>
<td>7B</td>
<td>Medical</td>
<td>10.13</td>
<td>5.23</td>
<td>19.76</td>
<td>5.88</td>
<td>18.16</td>
<td>6.63</td>
<td>22.94</td>
<td>8.25</td>
<td>16.12</td>
</tr>
<tr>
<td>7</td>
<td>Lingshu</td>
<td>7B</td>
<td>Medical</td>
<td>8.19</td>
<td>3.73</td>
<td>16.48</td>
<td>4.03</td>
<td>15.72</td>
<td>6.27</td>
<td>19.77</td>
<td>6.34</td>
<td>13.31</td>
</tr>
<tr>
<td>8</td>
<td>MedR1</td>
<td>2B</td>
<td>Medical</td>
<td>8.18</td>
<td>3.60</td>
<td>14.73</td>
<td>3.53</td>
<td>12.55</td>
<td>3.53</td>
<td>25.58</td>
<td>4.39</td>
<td>13.57</td>
</tr>
<tr>
<td>9</td>
<td>SAM4MLLM</td>
<td>8B</td>
<td>Grounding</td>
<td>7.94</td>
<td>—</td>
<td>16.49</td>
<td>6.30</td>
<td>14.69</td>
<td>5.81</td>
<td>12.61</td>
<td>6.24</td>
<td>11.96</td>
</tr>
<tr>
<td>10</td>
<td>Gemini-2.5-flash</td>
<td>—</td>
<td>General</td>
<td>7.86</td>
<td>3.24</td>
<td>14.29</td>
<td>3.99</td>
<td>5.69</td>
<td>7.77</td>
<td>16.37</td>
<td>7.15</td>
<td>13.91</td>
</tr>
<tr>
<td>11</td>
<td>Chiron-o1</td>
<td>8B</td>
<td>Medical</td>
<td>6.40</td>
<td>2.46</td>
<td>10.05</td>
<td>3.82</td>
<td>6.90</td>
<td>4.20</td>
<td>12.86</td>
<td>5.53</td>
<td>11.31</td>
</tr>
<tr>
<td>12</td>
<td>InternVL3</td>
<td>8B</td>
<td>General</td>
<td>5.70</td>
<td>2.46</td>
<td>9.23</td>
<td>3.72</td>
<td>6.54</td>
<td>3.67</td>
<td>14.44</td>
<td>3.78</td>
<td>8.71</td>
</tr>
<tr>
<td>13</td>
<td>MedGamma</td>
<td>4B</td>
<td>Medical</td>
<td>5.39</td>
<td>1.90</td>
<td>8.90</td>
<td>4.23</td>
<td>6.92</td>
<td>3.41</td>
<td>4.78</td>
<td>3.17</td>
<td>3.90</td>
</tr>
<tr>
<td>14</td>
<td>InternVL3</td>
<td>78B</td>
<td>General</td>
<td>4.02</td>
<td>1.55</td>
<td>7.23</td>
<td>2.04</td>
<td>2.95</td>
<td>2.12</td>
<td>12.21</td>
<td>1.33</td>
<td>8.19</td>
</tr>
<tr>
<td>15</td>
<td>MiniInternVL</td>
<td>4B</td>
<td>Medical</td>
<td>2.88</td>
<td>0.85</td>
<td>4.76</td>
<td>1.88</td>
<td>2.67</td>
<td>1.60</td>
<td>7.99</td>
<td>1.56</td>
<td>3.76</td>
</tr>
<tr>
<td>16</td>
<td>GPT-4o</td>
<td>—</td>
<td>General</td>
<td>2.65</td>
<td>1.12</td>
<td>4.72</td>
<td>0.92</td>
<td>0.91</td>
<td>0.36</td>
<td>11.70</td>
<td>1.01</td>
<td>4.16</td>
</tr>
</tbody>
</table>
</div>
</div>
</div>
</div>
</section>
<!----------------------------------Case Studies---------------------------------->
<section class="hero is-light is-small">
<div class="hero-body has-text-centered">
<h1 class="title is-1">🔡 Case Studies</h1>
</div>
</section>
<section>
<div class="container">
<div class="columns is-centered has-text-centered m-6">
<div class="column is-four-fifths">
<div class="content has-text-justified">
<h2 class="title is-4 has-text-centered mb-4">
🧩 Meta Information Examples
</h2>
<div id="case-carousel" class="carousel results-carousel">
<div class="box m-5" style="height: 100%; display: flex; justify-content: center; align-items: center; text-align: center;">
<div class="content has-text-centered">
<img src="static/images/meta_infor_organ1.png" width="90%"/>
<img src="static/images/meta_infor_organ2.png" width="90%"/>
<img src="static/images/meta_infor_organ4.png" width="90%"/>
<img src="static/images/meta_infor_organ5.png" width="90%"/>
<p>
Examples of the meta information from four anatomical super-categories. From top to bottom: <b>Abdomen</b>, <b>Abdomen</b>, <b>Brain</b>, <b>Lung</b> and <b>Heart</b>. Each case presents the image along with its: <i>modal</i>, <i>health</i>, <i>super-category</i>, <i>category</i>, <i>short description</i> and <i>long description</i>.
</p>
</div>
</div>
<div class="box m-5" style="height: 100%; display: flex; justify-content: center; align-items: center; text-align: center;">
<div class="content has-text-centered">
<img src="static/images/meta_infor_abnorm1.png" width="90%"/>
<img src="static/images/meta_infor_abnorm2.png" width="90%"/>
<p>
Examples of the meta information from two lesions super-categories. From top to bottom: <b>Neoplasm</b> and <b>Non-Neoplasm</b>. Each case presents the image along with its: <i>modal</i>, <i>health</i>, <i>super-category</i>, <i>category</i>, <i>short description</i> and <i>long description</i>.
</p>
</div>
</div>
</div>
</div>
</div>
</div>
</div>
</section>
<section>
<div class="container">
<div class="columns is-centered has-text-centered m-6">
<div class="column is-four-fifths">
<div class="content has-text-justified">
<h2 class="title is-4 has-text-centered mb-4">
🧩 QA Pairs Examples
</h2>
<div id="case-carousel" class="carousel results-carousel">
<div class="box m-5" style="height: 100%; display: flex; justify-content: center; align-items: center; text-align: center;">
<div class="content has-text-centered">
<img src="static/images/qa_pairs_organ1.png" width="80%"/>
<img src="static/images/qa_pairs_organ2.png" width="80%"/>
<img src="static/images/qa_pairs_organ4.png" width="80%"/>
<img src="static/images/qa_pairs_organ5.png" width="80%"/>
<p>
Examples of the qa pairs from three anatomical super-categories. From top to bottom: <b>Abdomen</b>, <b>Abdomen</b>, <b>Brain</b>, <b>Lung</b> and <b>Heart</b>. Each case presents the image along with its: <i>modal</i>, <i>health</i>, <i>super-category</i>, <i>category</i>, <i>question</i>, <i>think</i>, and <i>answer (bbox, point)</i>.
</p>
</div>
</div>
<div class="box m-5" style="height: 100%; display: flex; justify-content: center; align-items: center; text-align: center;">
<div class="content has-text-centered">
<img src="static/images/qa_pairs_abnorm1.png" width="80%"/>
<img src="static/images/qa_pairs_abnorm2.png" width="80%"/>
<p>
Examples of the qa pairs from two lesions super-categories. From top to bottom: <b>Neoplasm</b> and <b>Non-Neoplasm</b>. Each case presents the image along with its: <i>modal</i>, <i>health</i>, <i>super-category</i>, <i>category</i>, <i>question</i>, <i>think</i>, and <i>answer (bbox, point)</i>.
</p>
</div>
</div>
</div>
</div>
</div>
</div>
</div>
</section>
<!----------------------------------Citation---------------------------------->
<section class="section" id="BibTeX">
<div class="container is-max-desktop content">
<h2 class="title">📚 BibTeX Citation</h2>
<pre><code>
@article{yan2025medreasoner,
title={MedReasoner: Reinforcement Learning Drives Reasoning Grounding from Clinical Thought to Pixel-Level Precision},
author={Yan, Zhonghao and Diao, Muxi and Yang, Yuxuan and Xu, Jiayuan and Zhang, Kaizhou and Jing, Ruoyan and Yang, Lele and Liu, Yanxi and Liang, Kongming and Ma, Zhanyu},
journal={arXiv preprint arXiv:2508.08177},
year={2025}
}
</code></pre>
</div>
</section>
<!--footer-->
<footer class="footer">
<div class="columns is-centered">
<div class="column is-8">
<div class="content">
<p>
This website is adapted from <a href="https://we-math.github.io/">We-Math</a> and <a href="https://csbench.github.io/">CS-Bench</a>, licensed under a <a rel="license"
href="http://creativecommons.org/licenses/by-sa/4.0/">Creative
Commons Attribution-ShareAlike 4.0 International License</a>.
</p>
</div>
</div>
</div>
</footer>
<!----------------------------------End Body---------------------------------->
</body>
</html>