PathXAI: Evidence-guided structured pathology reporting by transferable tissue segmentation
Organizations: Intelligent Digital and Molecular Pathology Lab, Bioinformatics Institute, A∗STAR, Singapore. · Department of Pathology, The First Affiliated Hospital, Zhejiang University School of Medicine, Hangzhou, China. · Artificial Intelligence Innovation and Incubation Institute, Fudan University, Shanghai, China. · School of Big Data and Software Engineering, Chongqing University, Chongqing, China. ·
Abstract
Pathology reporting translates tissue observations across regions and slides into diagnoses and supporting morphological descriptions, yet explicit tissue-composition measurements are rarely used to guide H&E-based report generation. We developed PathXAI, which pairs transferable tissue segmentation with a pretrained vision-language model. Because supplying tissue information directly with H&E images gave limited benefit, PathXAI-Report uses segmentation overlays and tissue statistics in separate steps, selects between candidate diagnoses by fixed rules, acquires further images when evidence is insufficient and records every step for review, without updating model weights. Its segmentation component exceeded baselines in mean Dice on internal and external tests. In automated evaluation on 1,200 public and 151 private cases, PathXAI achieved higher agreement with reference diagnoses and documented morphology than H&E-based comparators. On the public cohort, it exceeded Qwen+H&E by 4.4 percentage points in cancer family match rate and 5.7 percentage points in finding F1.
Figures & tables
| Source | Conversion | Supervised classes |
|---|---|---|
| Pixel-labelled sources | ||
| BCSS | Source IDs ; ; ; ; ; ; . Remapped zero pixels are excluded from supervised Dice/BCE and the known-background penalty. | 1–6 |
| GlaS | Annotated gland objects ; remaining pixels . The source train, test-a and test-b images are pooled before the study partition. | 6 |
| DigestPath2019 | Grayscale mask , otherwise . Only paired positive images with nonempty foreground are retained; the negative-image directory is excluded. | 1 |
| WSSS4LUAD | Source IDs , , ; other values . Images and masks come from the source validation subset before the study partition. | 1, 3, 4 |
| BACH whole slides | Preprocessed integer mask IDs are retained without further remapping. Consistency with the XML-to-mask conversion script has not been verified. | 1, 4 |
| Setting | MedSAM3 | SAM2 |
|---|---|---|
| Initial learning rates | Supervised adaptation: visual adapters and class embeddings, . Joint supervision and focal refinement: visual adapters and configured auxiliary-parameter group, ; class embeddings, 0. | Initial fine-tuning: visual adapters and decoder, . Subsequent supervised adaptation: ; joint supervision and focal refinement: . Auxiliary-head rate, . |
| Cosine schedule | Adaptation: horizon 5,000 optimizer updates, clipped at 1% of each initial rate. Continuation: horizon 2,500 optimizer updates, minimum 0. | For configured microsteps, horizon optimizer updates; minimum rate is 1% of the smaller initial adapter and decoder rates. |
| Gradient accumulation | Loader positions count all samples, including skipped samples with no loss. At positions 8, 16, that contribute a loss, accumulated gradients are updated. Losses are not divided by eight. At epoch end, remaining gradients are updated if the loss-bearing sample count is not divisible by eight. | Eight microsteps per update; each loss is divided by eight. A final partial window is updated at the last microstep. |
| Scheduler updates | Once per optimizer-update boundary. The training iteration count advances per loss-bearing sample, separately from the scheduler counter. | Once per optimizer update, including the final partial window. |
| Mixed precision | BF16 autocast with gradient scaling. | CUDA autocast with gradient scaling. |
| Gradient clipping | Segmentation and classification parameter groups clipped separately to norm 1. | Decoder plus visual adapters, and auxiliary heads, clipped separately to norm 1. |
| Site | Abbrev. | Hospital | Cases |
|---|---|---|---|
| Breast | TTSH | Tan Tock Seng Hospital | 76 |
| Prostate | ZSH | Zhongshan Hospital, Fudan University | 12 |
| Prostate | CHH | Changhai Hospital | 12 |
| Prostate | CZH | Shanghai Changzheng Hospital | 12 |
| Prostate | HBGH | Hebei General Hospital | 12 |
| Prostate | JPH | Jiangsu Province Hospital | 14 |
| Characteristic | PathXAI-Report pipeline development subset | Evaluation cohort |
|---|---|---|
| Cases | 456 | 1,200 |
| Age at GDC index (years) | ||
| Available | 453 (99.3%) | 1,198 (99.8%) |
| Median (IQR) | 59.0 (50.0–67.0) | 58.0 (47.0–68.0) |
| Missing | 3 (0.7%) | 2 (0.2%) |
| Age privacy flag | 5 (1.1%) | 12 (1.0%) |
| Method | Image evidence | Additional context and output |
|---|---|---|
| Qwen+H&E; MedGemma | Selected H&E ROIs. | Known site; generated report. |
| DeepSeek+H&E; GLM+H&E | One to three fixed initial H&E ROIs, shared across these models; maximum image edge 768 pixels. | Known site; diagnostic call followed by report generation. |
| PLIP-guided Qwen | PLIP ranks and classifies three central 224-by-224-pixel crops from saved ROIs. | Crop descriptions and known site enter Qwen; single-pass report generation. |
| PRISM2 | One selected WSI per case, preferentially a diagnostic slide; WSI features computed by the released model. | Known site; one diagnostic line, without grade or morphology reporting. |
| HistGen | Features from all listed slides, concatenated per case. | No separate site prompt; native report decoder. |
| HistoGPT | Features from all listed slides, concatenated per case. | “Final diagnosis:” prompt without a separate site field; native report generation without GPT refinement. |
| Component | Compared configurations | Changed inputs or rules | Cases and results |
|---|---|---|---|
| Public evaluation cohort (1,200 cases) | |||
| Direct tissue input | H&E+tissue statistics versus H&E alone; H&E+overlays versus H&E alone | Supplies numeric tissue summaries or segmentation overlays for direct reporting, without KEEP retrieval or reinspection. | 1,200; Section 1.1.4 |
| Complete workflow | PathXAI-Report versus direct tissue-statistics input; PathXAI-Report versus direct overlay input | Adds diagnostic review, reinspection and constrained writing to tissue evidence. These are comparisons of complete workflows. | 1,200; Sections 1.1.4 and 1.1.4 |
| Grade-focused images | Report-only grade evidence versus PathXAI-Report (core) | Adds auxiliary H&E regions to writing only; diagnostic candidates do not receive these regions. | 1,200; 4.3. Report-workflow comparisons |
| Knowledge and diagnostic constraints | PathXAI-Report versus report-only grade evidence | Adds KEEP routing, invasion-state selection safeguards and diagnosis-locked writing. The comparison evaluates these changes together, including any resulting changes in acquired evidence. | 1,200; 4.3. Report-workflow comparisons |
| PathXAI-Report pipeline development subset (456 cases) | |||
| Measure | What it asks | Details |
| Five measures shown in the automated report evaluation figure | ||
| Cancer family match rate | What fraction of diagnoses belong to the target cancer family? Public targets use TCGA labels; private targets use each pathologist’s definite cancer diagnosis. Benign or unresolved private references are inapplicable. | Methods |
| Exact diagnosis match rate | What fraction of diagnoses match the specific entity or subtype in the case’s reference report? Synonyms count as matches; related subtypes do not. A definite, specific reference diagnosis is required. | Methods |
| Finding F1 | How well does the report balance the correctness of its finding statements and recovery of documented reference findings? It combines precision and recall; matching presence and matching absence both count. Undocumented finding types are not scored. | Methods |
| Finding recall | What fraction of documented reference finding types is correctly reproduced as present or absent? Missing, uncertain, conflicting and opposite statements receive no credit. | Methods |
| Finding contradiction rate | Among documented finding types that the report definitely calls present or absent, what fraction disagrees with the reference? Undefined when none is addressed; interpreted alongside coverage because omissions are not contradictions. | Methods |
| Term | Definition |
|---|---|
| System and modules | |
| PathXAI | The integrated study system, comprising PathXAI-Seg and PathXAI-Report. |
| Structured report with tissue evidence | A structured pathology report with diagnostic and descriptive sections and recorded associations to slide, ROI, tissue and contextual evidence. |
| PathXAI-Seg | The tissue-segmentation module. It produces predicted tissue label maps, aligned segmentation overlays, and segmentation-derived tissue statistics. |
| PathXAI-Report | The complete report-generation module, including evidence acquisition, diagnostic reasoning, and structured report writing. |
| PathXAI-Report (core) | The report-module comparator with the core dual-path workflow. Relative to the complete configuration, it lacks KEEP retrieval and auxiliary grade-focused writer inputs and uses a different candidate-selection policy without diagnosis locking. |
| Method | Cancer family match rate | Exact diagnosis match rate | Finding F1 | Finding recall | Finding contradiction rate |
|---|---|---|---|---|---|
| Qwen+H&E | 68.1 [65.36, 70.72] n=1200 | 58.9 [54.90, 62.89] n=604 | 43.0 [40.63, 45.42] n=997 | 36.6 [34.34, 38.88] n=997 | 16.0 [13.72, 18.42] n=707 |
| DeepSeek+H&E | 71.9 [69.28, 74.44] n=1200 | 60.3 [56.24, 64.19] n=604 | 42.1 [39.82, 44.43] n=997 | 36.1 [33.98, 38.40] n=997 | 23.2 [20.57, 25.84] n=738 |
| GLM+H&E | 61.8 [59.02, 64.59] n=1200 | 56.0 [51.90, 59.96] n=604 | 44.2 [41.85, 46.47] n=997 | 38.1 [35.86, 40.42] n=997 | 22.5 [19.96, 25.22] n=751 |
| MedGemma | 6.7 [5.32, 8.23] n=1200 | 6.1 [4.35, 8.35] n=604 | 37.4 [35.12, 39.70] n=997 | 31.8 [29.54, 33.92] n=997 | 14.4 [11.90, 17.02] n=610 |
| PLIP-guided Qwen | 45.7 [42.82, 48.53] n=1200 | 39.2 [35.32, 43.26] n=604 | 21.9 [19.84, 24.04] n=997 | 17.8 [15.97, 19.65] n=997 | 6.3 [3.99, 8.83] n=351 |
| PRISM2 | 53.2 [50.38, 56.10] n=1200 | 52.6 [48.58, 56.69] n=604 | NA | NA | NA |
| Method | Cancer family match rate | Exact diagnosis match rate | Finding F1 | Finding recall | Finding contradiction rate |
|---|---|---|---|---|---|
| HistGen | 39.1 [32.51, 45.94] n=215 | 42.5 [30.97, 54.59] n=73 | 17.2 [13.13, 21.54] n=173 | 13.6 [10.12, 17.33] n=173 | 32.2 [22.97, 41.67] n=74 |
| HistoGPT | 5.1 [2.58, 8.97] n=215 | 1.4 [0.03, 7.40] n=73 | 12.0 [8.43, 15.73] n=173 | 8.8 [6.15, 11.84] n=173 | 9.3 [2.33, 18.60] n=43 |
| PathXAI-Report (core) | 73.5 [67.06, 79.26] n=215 | 58.9 [46.77, 70.29] n=73 | 43.2 [38.01, 48.50] n=173 | 35.8 [31.06, 40.86] n=173 | 12.5 [8.13, 17.66] n=126 |
| PathXAI-Report | 77.7 [71.51, 83.06] n=215 | 63.0 [50.91, 74.03] n=73 | 43.3 [38.01, 48.68] n=173 | 36.0 [31.12, 41.04] n=173 | 12.7 [8.12, 17.87] n=127 |
| Method | Cancer family match rate | Exact diagnosis match rate | Finding F1 | Finding recall | Finding contradiction rate |
|---|---|---|---|---|---|
| Qwen+H&E | +4.42 [1.92, 6.75] n=1200 | +3.31 [0.00, 6.62] n=604 | +5.75 [3.63, 7.87] n=997 | +5.07 [3.04, 7.03] n=997 | -5.52 [-8.02, -3.11] n=664 |
| DeepSeek+H&E | +0.58 [-1.92, 3.17] n=1200 | +1.99 [-1.82, 5.79] n=604 | +6.67 [4.29, 9.27] n=997 | +5.49 [3.11, 7.78] n=997 | -9.19 [-12.11, -6.17] n=634 |
| GLM+H&E | +10.67 [7.42, 13.75] n=1200 | +6.29 [2.15, 10.60] n=604 | +4.60 [1.89, 7.15] n=997 | +3.50 [1.06, 5.96] n=997 | -10.01 [-13.07, -7.04] n=644 |
| MedGemma | +65.83 [63.17, 68.42] n=1200 | +56.13 [52.15, 60.10] n=604 | +11.33 [8.75, 13.89] n=997 | +9.86 [7.46, 12.35] n=997 | -1.08 [-3.97, 1.75] n=546 |
| PLIP-guided Qwen | +26.83 [24.08, 29.50] n=1200 | +23.01 [19.21, 26.66] n=604 | +26.88 [24.32, 29.34] n=997 | +23.87 [21.53, 26.14] n=997 | +4.75 [1.92, 7.43] n=327 |
| PRISM2 | +19.25 [15.50, 22.84] n=1200 | +9.60 [4.47, 14.74] n=604 | NA | NA | NA |
| Method | PathXAI-Report pipeline development match rate (%) | Evaluation match rate (%) | Evaluation difference (pp) [95% CI] |
|---|---|---|---|
| Constant diagnosis | 14.9 | 32.8 | — |
| Qwen+H&E | 71.5 | 68.1 | +35.33 [+32.17, +38.50] |
| DeepSeek+H&E | 70.6 | 71.9 | +39.17 [+36.17, +42.17] |
| GLM+H&E | 55.9 | 61.8 | +29.08 [+25.67, +32.42] |
| MedGemma | 7.7 | 6.7 | -26.08 [-29.00, -22.92] |
| PLIP-guided Qwen | 29.8 | 45.7 | +12.92 [+9.33, +16.58] |
| Metric | Qwen+H&E | DeepSeek+H&E | GLM+H&E |
|---|---|---|---|
| Diagnostic composite (cohort target) | 84.1 [82.9, 85.2] n=1656 | 83.9 [82.8, 85.2] n=1656 | 73.2 [71.5, 74.8] n=1656 |
| Cancer family match rate | 69.0 [66.7, 71.2] n=1656 | 71.6 [69.3, 73.7] n=1656 | 60.2 [57.8, 62.6] n=1656 |
| Broad diagnostic agreement | 96.7 [95.7, 97.5] n=1656 | 94.4 [93.2, 95.4] n=1656 | 83.4 [81.5, 85.2] n=1656 |
| Malignancy agreement | 90.0 [88.4, 91.4] n=1656 | 88.4 [86.8, 89.9] n=1656 | 79.5 [77.4, 81.4] n=1656 |
| Invasion agreement | 69.7 [66.6, 72.7] n=894 | 59.4 [56.1, 62.6] n=894 | 61.9 [58.6, 65.1] n=894 |
| Necrosis agreement | 46.2 [41.4, 51.1] n=424 | 46.0 [41.2, 50.9] n=424 | 37.0 [32.4, 41.8] n=424 |
| Metric | MedGemma | PLIP-guided Qwen | PRISM2 |
|---|---|---|---|
| Diagnostic composite (cohort target) | 41.9 [40.5, 43.5] n=1656 | 70.7 [69.3, 72.1] n=1656 | 74.9 [73.5, 76.4] n=1656 |
| Cancer family match rate | 6.9 [5.8, 8.3] n=1656 | 41.3 [38.9, 43.7] n=1656 | 54.0 [51.5, 56.4] n=1656 |
| Broad diagnostic agreement | 66.3 [64.0, 68.6] n=1656 | 89.7 [88.1, 91.1] n=1656 | 90.5 [89.0, 91.9] n=1656 |
| Malignancy agreement | 66.8 [64.5, 69.1] n=1656 | 95.0 [93.8, 96.0] n=1656 | 87.4 [85.7, 89.0] n=1656 |
| Invasion agreement | 56.9 [53.6, 60.2] n=894 | 22.3 [19.6, 25.1] n=894 | NA |
| Necrosis agreement | 54.5 [49.6, 59.3] n=424 | 14.4 [11.2, 18.1] n=424 | NA |
| Metric | PathXAI-Report (core) | PathXAI-Report |
|---|---|---|
| Diagnostic composite (cohort target) | 84.4 [83.3, 85.5] n=1656 | 86.6 [85.7, 87.6] n=1656 |
| Cancer family match rate | 69.7 [67.4, 71.9] n=1656 | 73.7 [71.5, 75.8] n=1656 |
| Broad diagnostic agreement | 96.3 [95.2, 97.1] n=1656 | 97.6 [96.8, 98.3] n=1656 |
| Malignancy agreement | 90.8 [89.3, 92.2] n=1656 | 91.2 [89.7, 92.5] n=1656 |
| Invasion agreement | 70.7 [67.6, 73.7] n=894 | 77.4 [74.5, 80.1] n=894 |
| Necrosis agreement | 69.8 [65.2, 74.1] n=424 | 67.9 [63.3, 72.3] n=424 |
| Method | PathXAI-Report pipeline development subset (456) Exact diagnosis match rate | Evaluation cohort (1,200) Exact diagnosis match rate | PathXAI-Report pipeline development subset (456) Diagnostic composite | Evaluation cohort (1,200) Diagnostic composite |
|---|---|---|---|---|
| Qwen+H&E | 61.6 [56.3, 66.8] n=352 | 58.9 [54.9, 62.9] n=604 | 81.0 [78.4, 83.5] n=456 | 88.6 [87.5, 89.7] n=1200 |
| DeepSeek+H&E | 62.5 [57.2, 67.6] n=352 | 60.3 [56.2, 64.2] n=604 | 81.5 [79.0, 83.9] n=456 | 86.3 [84.9, 87.7] n=1200 |
| GLM+H&E | 46.6 [41.3, 52.0] n=352 | 56.0 [51.9, 60.0] n=604 | 69.1 [65.9, 72.4] n=456 | 77.5 [75.5, 79.4] n=1200 |
| MedGemma | 3.7 [2.0, 6.2] n=352 | 6.1 [4.3, 8.3] n=604 | 43.1 [40.0, 46.3] n=456 | 54.7 [52.5, 57.0] n=1200 |
| PLIP-guided Qwen | 23.9 [19.5, 28.7] n=352 | 39.2 [35.3, 43.3] n=604 | 56.6 [53.4, 59.9] n=456 | 86.0 [84.8, 87.2] n=1200 |
| PRISM2 | 49.1 [43.8, 54.5] n=352 | 52.6 [48.6, 56.7] n=604 | 73.2 [70.1, 76.3] n=456 | 83.5 [82.0, 85.1] n=1200 |
| System | Method | Balanced grade accuracy [95% CI] | n | K | Always-high [1/K] | Case-average exact grade agreement (sensitivity) | Grade macro-F1 (sensitivity) |
|---|---|---|---|---|---|---|---|
| Fuhrman (1–4) | DeepSeek+H&E | 0.0 [0.0, 27.4] n=118 | 118 | 4 | 25.0 [25.0] | 0.0 | 0.0 |
| Fuhrman (1–4) | GLM+H&E | 0.0 [0.0, 27.4] n=118 | 118 | 4 | 25.0 [25.0] | 0.0 | 0.0 |
| Fuhrman (1–4) | PLIP-guided Qwen | 0.0 [0.0, 27.4] n=118 | 118 | 4 | 25.0 [25.0] | 0.0 | 0.0 |
| Fuhrman (1–4) | PathXAI-Report (core) | 0.0 [0.0, 27.4] n=118 | 118 | 4 | 25.0 [25.0] | 0.0 | 0.0 |
| Fuhrman (1–4) | PathXAI-Report | 0.0 [0.0, 27.4] n=118 | 118 | 4 | 25.0 [25.0] | 0.0 | 0.0 |
| Fuhrman (1–4) | MedGemma | 0.0 [0.0, 27.4] n=118 | 118 | 4 | 25.0 [25.0] | 0.0 | 0.0 |
| System | Method | Ordinal grade agreement [95% CI] |
|---|---|---|
| Fuhrman (1–4) | DeepSeek+H&E | 0.0 [0.0, 4.0] n=118 |
| Fuhrman (1–4) | GLM+H&E | 0.0 [0.0, 4.0] n=118 |
| Fuhrman (1–4) | PLIP-guided Qwen | 0.0 [0.0, 4.0] n=118 |
| Fuhrman (1–4) | PathXAI-Report (core) | 0.0 [0.0, 4.0] n=118 |
| Fuhrman (1–4) | PathXAI-Report | 0.0 [0.0, 4.0] n=118 |
| Fuhrman (1–4) | MedGemma | 0.0 [0.0, 4.0] n=118 |
| Comparator | Metric | n | Difference [95% CI] |
|---|---|---|---|
| Qwen+H&E | Diagnostic composite (cohort target) | 1656 | 2.5 [1.6, 3.5] |
| Qwen+H&E | Cancer family match rate | 1656 | 4.6 [2.7, 6.6] |
| Qwen+H&E | Finding F1 | 1405 | 5.5 [3.7, 7.3] |
| Qwen+H&E | Finding coverage (A/D) | 1405 | 4.1 [2.5, 5.8] |
| Qwen+H&E | Finding recall (C/D) | 1405 | 4.8 [3.2, 6.5] |
| Qwen+H&E | Finding contradiction rate (E/A) | 938 | -5.8 [-7.9, -3.8] |
| Method | Overall accuracy correct/N (%) | Coverage n/N (%) | Conditional accuracy correct/n (%) | Balanced accuracy (%) |
|---|---|---|---|---|
| PathXAI-Report pipeline development subset (applicable n=168) | ||||
| Qwen+H&E | 50/168 (29.8) | 61/168 (36.3) | 50/61 (82.0) | 24.7 |
| DeepSeek+H&E | 104/168 (61.9) | 144/168 (85.7) | 104/144 (72.2) | 49.7 |
| GLM+H&E | 90/168 (53.6) | 130/168 (77.4) | 90/130 (69.2) | 39.8 |
| MedGemma | 54/168 (32.1) | 91/168 (54.2) | 54/91 (59.3) | 18.9 |
| PLIP-guided Qwen | 82/168 (48.8) | 145/168 (86.3) | 82/145 (56.6) | 28.5 |
| Method | Answered / eligible | Overall accuracy (%) | Weighted [95% CI] |
|---|---|---|---|
| Public evaluation cohort | |||
| Qwen+H&E | 222 / 472 | 30.9 | 0.387 [0.288, 0.486] |
| DeepSeek+H&E | 379 / 472 | 48.5 | 0.298 [0.220, 0.376] |
| GLM+H&E | 347 / 472 | 47.9 | 0.415 [0.328, 0.500] |
| MedGemma | 278 / 472 | 27.8 | -0.021 [-0.062, 0.025] |
| PLIP-guided Qwen | 471 / 472 | 46.6 | 0.000 [0.000, 0.000] |
| Evaluation set | Before exclusion | Excluded | Evaluated |
| Public benchmark | 1,708 | 52 | 1,656 |
| PathXAI-Report pipeline development subset | 508 | 52 | 456 |
| Public evaluation cohort | 1,200 | 0 | 1,200 |
| Four reporting policies | 1,200 | 0 | 1,200 |
| Pathologist ranking and correction assessment | 145 | 13 | 132 |
| HistGen/HistoGPT four-method comparison | 324 | 12 | 312 |
| Comparator | Paired before / after | Primary difference (pp) | After exclusion (pp) [95% CI] |
|---|---|---|---|
| Cancer family match rate | |||
| Qwen+H&E | 1200 / 1175 | +4.42 | +4.60 [+2.21, +7.06] |
| DeepSeek+H&E | 1200 / 1175 | +0.58 | +0.68 [-1.96, +3.32] |
| GLM+H&E | 1200 / 1175 | +10.67 | +10.30 [+7.06, +13.53] |
| MedGemma | 1200 / 1175 | +65.83 | +65.70 [+62.98, +68.43] |
| PLIP-guided Qwen | 1200 / 1175 | +26.83 | +27.32 [+24.60, +30.04] |
| Method | n ( ) | n (A=0) | Finding F1 | Finding coverage (A/D) | Finding recall (C/D) | Finding contradiction rate (E/A) |
|---|---|---|---|---|---|---|
| Qwen+H&E | 1405 | 401 | 42.3 [40.4, 44.3] n=1405 | 42.5 [40.6, 44.3] n=1405 | 35.7 [33.9, 37.6] n=1405 | 17.0 [15.0, 19.0] n=1004 |
| DeepSeek+H&E | 1405 | 399 | 39.7 [37.8, 41.8] n=1405 | 44.8 [42.8, 46.7] n=1405 | 33.8 [32.1, 35.6] n=1405 | 24.1 [21.8, 26.5] n=1006 |
| GLM+H&E | 1405 | 352 | 43.2 [41.3, 45.1] n=1405 | 48.0 [46.1, 49.9] n=1405 | 37.1 [35.3, 39.0] n=1405 | 23.1 [20.9, 25.3] n=1053 |
| MedGemma | 1405 | 566 | 35.6 [33.6, 37.5] n=1405 | 35.3 [33.4, 37.2] n=1405 | 29.9 [28.0, 31.8] n=1405 | 15.3 [13.1, 17.6] n=839 |
| PLIP-guided Qwen | 1405 | 933 | 20.1 [18.5, 21.8] n=1405 | 17.1 [15.6, 18.7] n=1405 | 16.1 [14.7, 17.6] n=1405 | 7.2 [5.1, 9.4] n=472 |
| PathXAI-Report (core) | 1405 | 399 | 45.2 [43.2, 47.2] n=1405 | 42.9 [41.1, 44.8] n=1405 | 37.9 [36.0, 39.7] n=1405 | 11.1 [9.4, 12.8] n=1006 |
| Method | F1 | Coverage ( ) | Recall ( ) | Contradictions ( ) | Pooled |
|---|---|---|---|---|---|
| Public evaluation cohort (n=997 cases with documented findings) | |||||
| Qwen+H&E | 43.0 | 42.9 | 36.6 | 6.3 [5.3, 7.4] | 7.1 |
| DeepSeek+H&E | 42.1 | 47.5 | 36.1 | 11.3 [9.8, 12.8] | 10.8 |
| GLM+H&E | 44.2 | 48.6 | 38.1 | 10.5 [9.1, 11.9] | 11.1 |
| MedGemma | 37.4 | 37.0 | 31.8 | 5.2 [4.3, 6.2] | 5.0 |
| PLIP-guided Qwen | 21.9 | 18.7 | 17.8 | 1.0 [0.6, 1.4] | 1.0 |
| Comparator | Cases | Fixed-reference difference | Shared cases / items | Shared-finding difference |
|---|---|---|---|---|
| Public evaluation cohort | ||||
| Qwen+H&E | 997 | -0.8 [-2.0, 0.5] | 656 / 827 | -6.0 [-8.5, -3.6] |
| PathXAI-Report (core) | 997 | 1.0 [-0.1, 2.0] | 672 / 867 | 0.9 [-0.6, 2.4] |
| Private reference set P1 | ||||
| Qwen+H&E | 150 | 6.7 [1.7, 11.7] | 133 / 215 | 7.1 [2.3, 12.4] |
| PathXAI-Report (core) | 150 | -2.7 [-7.3, 1.7] | 140 / 229 | -3.2 [-7.1, 0.7] |
| Finding | Public evaluation | P1 | P2 |
|---|---|---|---|
| Invasion | 612 / 29 | 120 / 23 | 122 / 24 |
| Necrosis | 258 / 48 | 8 / 141 | 15 / 120 |
| Lymphovascular invasion | 187 / 132 | 0 / 0 | 0 / 0 |
| Perineural invasion | 91 / 56 | 0 / 0 | 0 / 0 |
| Metastasis | 309 / 44 | 0 / 0 | 0 / 0 |
| Pleomorphism | 52 / 0 | 0 / 0 | 0 / 0 |
| Metric | HistGen | HistoGPT | PathXAI-Report (core) | PathXAI-Report |
|---|---|---|---|---|
| Cancer family match rate | 36.2 [30.9, 41.8] n=312 | 4.2 [2.2, 7.0] n=312 | 74.4 [69.1, 79.1] n=312 | 79.5 [74.6, 83.8] n=312 |
| Finding F1 | 17.2 [13.6, 20.8] n=256 | 12.3 [9.2, 15.6] n=256 | 44.9 [40.4, 49.1] n=256 | 45.7 [41.2, 49.9] n=256 |
| Finding coverage (A/D) | 21.0 [17.2, 25.0] n=256 | 10.5 [7.7, 13.5] n=256 | 43.4 [39.0, 47.6] n=256 | 44.9 [40.5, 49.4] n=256 |
| Finding recall (C/D) | 13.7 [10.9, 17.0] n=256 | 9.6 [7.0, 12.5] n=256 | 37.2 [33.3, 41.3] n=256 | 38.2 [34.3, 42.5] n=256 |
| Finding contradiction rate (E/A) | 32.4 [24.1, 40.6] n=106 | 8.3 [1.7, 15.0] n=60 | 12.5 [8.7, 16.7] n=189 | 12.7 [9.0, 16.6] n=193 |
| Balanced grade accuracy (histological 1–3) | 23.5 [ 9.4 , 48.8 ] n= 76 | 0.0 [0.0, 20.6] n= 76 | 23.4 [ 12.6 , 48.3 ] n= 76 | 32.8 [ 17.7 , 57.3 ] n= 76 |
| Metric | HistGen | HistoGPT | PathXAI-Report (core) | PathXAI-Report |
|---|---|---|---|---|
| Balanced grade accuracy | 16.7 [0.1, 86.8 ] n= 8 | 0.0 [0.0, 83.9 ] n= 8 | 0.0 [0.0, 83.9 ] n= 8 | 0.0 [0.0, 83.9 ] n= 8 |
| Exact grade agreement (sensitivity) | 12.5 [ 0.3 , 52.7 ] n= 8 | 0.0 [0.0, 36.9 ] n= 8 | 0.0 [0.0, 36.9 ] n= 8 | 0.0 [0.0, 36.9 ] n= 8 |
| Ordinal grade agreement (sensitivity) | 20.8 [ 1.2 , 67.8 ] n= 8 | 0.0 [0.0, 45.0 ] n= 8 | 0.0 [0.0, 45.0 ] n= 8 | 0.0 [0.0, 45.0 ] n= 8 |
| Metric | HistGen | HistoGPT | PathXAI-Report (core) | PathXAI-Report |
|---|---|---|---|---|
| Balanced grade accuracy | 0.0 [0.0, 97.5] n=1 | 0.0 [0.0, 97.5] n=1 | 0.0 [0.0, 97.5] n=1 | 0.0 [0.0, 97.5] n=1 |
| Exact grade agreement (sensitivity) | 0.0 [0.0, 97.5] n=1 | 0.0 [0.0, 97.5] n=1 | 0.0 [0.0, 97.5] n=1 | 0.0 [0.0, 97.5] n=1 |
| Ordinal grade agreement (sensitivity) | 0.0 [0.0, 99.7] n=1 | 0.0 [0.0, 99.7] n=1 | 0.0 [0.0, 99.7] n=1 | 0.0 [0.0, 99.7] n=1 |
| Metric | HistGen | HistoGPT | PathXAI-Report (core) | PathXAI-Report |
|---|---|---|---|---|
| Exact grade agreement (sensitivity) | 26.3 [ 16.9 , 37.7 ] n= 76 | 0.0 [0.0, 4.7 ] n= 76 | 34.2 [ 23.7 , 46.0 ] n= 76 | 42.1 [ 30.9 , 54.0 ] n= 76 |
| Ordinal grade agreement (sensitivity) | 33.6 [ 22.0 , 46.8 ] n= 76 | 0.7 [0.0, 6.9 ] n= 76 | 40.1 [ 27.7 , 53.5 ] n= 76 | 51.3 [ 38.2 , 64.3 ] n= 76 |
| Metric | HistGen | HistoGPT | PathXAI-Report (core) | PathXAI-Report |
|---|---|---|---|---|
| Balanced grade accuracy | 0.0 [0.0, 97.5 ] n= 1 | 0.0 [0.0, 97.5 ] n= 1 | 0.0 [0.0, 97.5 ] n= 1 | 0.0 [0.0, 97.5 ] n= 1 |
| Exact grade agreement (sensitivity) | 0.0 [0.0, 97.5 ] n= 1 | 0.0 [0.0, 97.5 ] n= 1 | 0.0 [0.0, 97.5 ] n= 1 | 0.0 [0.0, 97.5 ] n= 1 |
| Ordinal grade agreement (sensitivity) | 0.0 [0.0, 99.2 ] n= 1 | 0.0 [0.0, 99.2 ] n= 1 | 0.0 [0.0, 99.2 ] n= 1 | 0.0 [0.0, 99.2 ] n= 1 |
| Metric | HistGen | HistoGPT | PathXAI-Report (core) | PathXAI-Report |
|---|---|---|---|---|
| Balanced grade accuracy | 50.0 [ 0.3 , 90.8 ] n= 8 | 0.0 [0.0, 90.5 ] n= 8 | 0.0 [0.0, 90.5 ] n= 8 | 0.0 [0.0, 90.5 ] n= 8 |
| Exact grade agreement (sensitivity) | 25.0 [ 3.2 , 65.1 ] n= 8 | 0.0 [0.0, 36.9 ] n= 8 | 0.0 [0.0, 36.9 ] n= 8 | 0.0 [0.0, 36.9 ] n= 8 |
| Ordinal grade agreement (sensitivity) | 25.0 [ 1.8 , 71.7 ] n= 8 | 0.0 [0.0, 45.0 ] n= 8 | 12.5 [ 1.9 , 57.0 ] n= 8 | 16.7 [ 3.8 , 59.5 ] n= 8 |
| Metric | Comparator | Difference [95% CI], pp | |
|---|---|---|---|
| Cancer family match rate | HistGen | 312 | +43.3 [ +36.5 , +50.3 ] |
| Cancer family match rate | HistoGPT | 312 | +75.3 [+70.5, +79.8] |
| Finding F1 | HistGen | 256 | +28.5 [+23.0, +33.8] |
| Finding F1 | HistoGPT | 256 | +33.4 [+28.7, +37.9] |
| Finding coverage (A/D) | HistGen | 256 | +23.9 [+18.8, +29.3] |
| Finding coverage (A/D) | HistoGPT | 256 | +34.4 [+29.9, +39.1] |
| Metric | HistGen | PathXAI-Report (core) | PathXAI-Report |
|---|---|---|---|
| Cancer family match rate | 36.0 [30.7, 41.6] n=314 | 74.5 [69.3, 79.2] n=314 | 79.3 [74.4, 83.6] n=314 |
| Finding F1 | 17.0 [13.8, 20.6] n=258 | 45.0 [40.5, 49.3] n=258 | 45.3 [41.0, 49.6] n=258 |
| Finding coverage (A/D) | 20.8 [16.9, 24.7] n=258 | 43.6 [39.3, 47.7] n=258 | 44.8 [40.4, 49.1] n=258 |
| Finding recall (C/D) | 13.6 [10.6, 16.8] n=258 | 37.3 [33.4, 41.3] n=258 | 37.9 [33.8, 42.1] n=258 |
| Finding contradiction rate (E/A) | 32.4 [24.1, 40.6] n=106 | 12.6 [8.9, 16.7] n=191 | 13.1 [9.2, 17.3] n=194 |
| Balanced grade accuracy (histological 1–3) | 22.7 [ 9.2 , 47.7] n= 77 | 23.0 [ 12.6 , 47.4 ] n= 77 | 32.0 [ 17.5 , 56.2 ] n= 77 |
| Metric | HistGen | PathXAI-Report (core) | PathXAI-Report |
|---|---|---|---|
| Balanced grade accuracy | 16.7 [0.1, 86.8 ] n= 8 | 0.0 [0.0, 83.9 ] n= 8 | 0.0 [0.0, 83.9 ] n= 8 |
| Exact grade agreement (sensitivity) | 12.5 [ 0.3 , 52.7 ] n= 8 | 0.0 [0.0, 36.9 ] n= 8 | 0.0 [0.0, 36.9 ] n= 8 |
| Ordinal grade agreement (sensitivity) | 20.8 [ 1.2 , 67.8 ] n= 8 | 0.0 [0.0, 45.0 ] n= 8 | 0.0 [0.0, 45.0 ] n= 8 |
| Metric | HistGen | PathXAI-Report (core) | PathXAI-Report |
|---|---|---|---|
| Balanced grade accuracy | 0.0 [0.0, 97.5] n=1 | 0.0 [0.0, 97.5] n=1 | 0.0 [0.0, 97.5] n=1 |
| Exact grade agreement (sensitivity) | 0.0 [0.0, 97.5] n=1 | 0.0 [0.0, 97.5] n=1 | 0.0 [0.0, 97.5] n=1 |
| Ordinal grade agreement (sensitivity) | 0.0 [0.0, 99.7] n=1 | 0.0 [0.0, 99.7] n=1 | 0.0 [0.0, 99.7] n=1 |
| Metric | HistGen | PathXAI-Report (core) | PathXAI-Report |
|---|---|---|---|
| Exact grade agreement (sensitivity) | 26.0 [ 16.6 , 37.2 ] n= 77 | 33.8 [ 23.4 , 45.4 ] n= 77 | 41.6 [ 30.4 , 53.4 ] n= 77 |
| Ordinal grade agreement (sensitivity) | 33.8 [ 22.3 , 46.9 ] n= 77 | 39.6 [ 27.3 , 52.9 ] n= 77 | 50.6 [ 37.7 , 63.6 ] n= 77 |
| Metric | HistGen | PathXAI-Report (core) | PathXAI-Report |
|---|---|---|---|
| Balanced grade accuracy | 0.0 [0.0, 97.5 ] n= 1 | 0.0 [0.0, 97.5 ] n= 1 | 0.0 [0.0, 97.5 ] n= 1 |
| Exact grade agreement (sensitivity) | 0.0 [0.0, 97.5 ] n= 1 | 0.0 [0.0, 97.5 ] n= 1 | 0.0 [0.0, 97.5 ] n= 1 |
| Ordinal grade agreement (sensitivity) | 0.0 [0.0, 99.2 ] n= 1 | 0.0 [0.0, 99.2 ] n= 1 | 0.0 [0.0, 99.2 ] n= 1 |
| Metric | HistGen | PathXAI-Report (core) | PathXAI-Report |
|---|---|---|---|
| Balanced grade accuracy | 50.0 [0.3, 89.0 ] n= 9 | 0.0 [0.0, 88.7 ] n= 9 | 0.0 [0.0, 88.7 ] n= 9 |
| Exact grade agreement (sensitivity) | 22.2 [ 2.8 , 60.0 ] n= 9 | 0.0 [0.0, 33.6 ] n= 9 | 0.0 [0.0, 33.6 ] n= 9 |
| Ordinal grade agreement (sensitivity) | 22.2 [ 1.6 , 66.6 ] n= 9 | 11.1 [ 1.7 , 52.8 ] n= 9 | 22.2 [ 5.5 , 62.3 ] n= 9 |
| Field | P1 | P2 | Comparable | Matching (%) | [95% CI] |
|---|---|---|---|---|---|
| Broad diagnosis | 123 | 123 | 123 | 123 (100.0%) | Undefined |
| Specific diagnosis | 136 | 137 | 127 | 122 (96.1%) | 0.939 [0.886, 0.987] |
| Malignancy | 145 | 149 | 145 | 143 (98.6%) | 0.951 [0.875, 1.000] |
| Invasion | 144 | 147 | 140 | 139 (99.3%) | 0.970 [0.892, 1.000] |
| Necrosis | 150 | 136 | 135 | 126 (93.3%) | 0.576 [0.283, 0.799] |
| Grade | 76 | 76 | 75 | 45 (60.0%) | 0.036 [-0.170, 0.246] |
| Method | P1 benign | P1 borderline | P2 benign | P2 borderline |
|---|---|---|---|---|
| Qwen+H&E | 11/13 | 4/9 | 11/13 | 6/11 |
| DeepSeek+H&E | 13/13 | 8/9 | 13/13 | 9/11 |
| GLM+H&E | 3/13 | 0/9 | 3/13 | 1/11 |
| MedGemma | 9/13 | 6/9 | 8/13 | 7/11 |
| PLIP-guided Qwen | 12/13 | 7/9 | 11/13 | 9/11 |
| PRISM2 | 11/13 | 0/9 | 11/13 | 0/11 |
| Method | Reference | Benign | Borderline | Malignant | Missing | Uncertain | Conflict |
|---|---|---|---|---|---|---|---|
| Reference category: Benign | |||||||
| Qwen+H&E | 13 / 13 | 1 / 1 | 0 / 0 | 11 / 11 | 1 / 1 | 0 / 0 | 0 / 0 |
| DeepSeek+H&E | 13 / 13 | 0 / 0 | 0 / 0 | 13 / 13 | 0 / 0 | 0 / 0 | 0 / 0 |
| GLM+H&E | 13 / 13 | 8 / 8 | 0 / 0 | 3 / 3 | 0 / 0 | 1 / 1 | 1 / 1 |
| MedGemma | 13 / 13 | 1 / 1 | 0 / 0 | 9 / 8 | 3 / 4 | 0 / 0 | 0 / 0 |
| PLIP-guided Qwen | 13 / 13 | 0 / 0 | 0 / 0 | 12 / 11 | 1 / 2 | 0 / 0 | 0 / 0 |
| Analysis | Method | Diagnostic composite | Exact diagnosis match rate | Finding F1 | Finding recall |
|---|---|---|---|---|---|
| Shared cases (151) / P1 | Qwen+H&E | 72.5 [66.6, 78.2] n=146 | 54.1 [45.3, 62.7] n=135 | 71.6 [65.9, 77.0] n=150 | 68.0 [62.3, 73.7] n=150 |
| Shared cases (151) / P1 | DeepSeek+H&E | 72.6 [66.3, 78.8] n=146 | 60.0 [51.2, 68.3] n=135 | 66.4 [61.2, 71.3] n=150 | 61.7 [56.3, 66.7] n=150 |
| Shared cases (151) / P1 | GLM+H&E | 61.9 [55.9, 67.9] n=146 | 43.0 [34.5, 51.8] n=135 | 63.1 [57.1, 68.8] n=150 | 56.7 [51.0, 62.7] n=150 |
| Shared cases (151) / P1 | MedGemma | 40.8 [35.8, 46.0] n=146 | 3.0 [0.8, 7.4] n=135 | 38.4 [33.0, 43.9] n=150 | 35.7 [30.3, 41.0] n=150 |
| Shared cases (151) / P1 | PLIP-guided Qwen | 59.5 [53.6, 65.5] n=146 | 32.6 [24.8, 41.2] n=135 | 31.1 [25.8, 36.4] n=150 | 23.3 [19.3, 27.3] n=150 |
| Shared cases (151) / P1 | PRISM2 | 74.4 [69.2, 79.6] n=146 | 54.8 [46.0, 63.4] n=135 | NA | NA |
| Reference | Method | Finding coverage (A/D) | Finding contradiction rate (E/A) |
|---|---|---|---|
| P1 | Qwen+H&E | 79.0 [73.7, 84.0] n=150 | 14.5 [9.8, 19.6] n=138 |
| P1 | DeepSeek+H&E | 78.7 [74.0, 83.3] n=150 | 19.7 [14.8, 24.6] n=142 |
| P1 | GLM+H&E | 63.3 [57.7, 69.0] n=150 | 10.9 [6.2, 16.0] n=128 |
| P1 | MedGemma | 63.0 [56.0, 69.3] n=150 | 41.6 [35.0, 48.2] n=113 |
| P1 | PLIP-guided Qwen | 30.7 [26.7, 34.7] n=150 | 23.1 [14.3, 31.9] n=91 |
| P1 | PRISM2 | NA | NA |
| Metric | Qwen+H&E | DeepSeek+H&E | GLM+H&E |
|---|---|---|---|
| Broad diagnostic agreement | 98.4 [94.2, 99.8] n=123 | 95.1 [89.7, 98.2] n=123 | 84.6 [76.9, 90.4] n=123 |
| Malignancy agreement | 84.0 [77.0, 89.6] n=144 | 81.2 [73.9, 87.3] n=144 | 74.3 [66.4, 81.2] n=144 |
| Invasion agreement | 67.8 [59.5, 75.4] n=143 | 76.9 [69.1, 83.6] n=143 | 67.1 [58.8, 74.8] n=143 |
| Necrosis agreement | 68.5 [60.3, 75.8] n=149 | 47.7 [39.4, 56.0] n=149 | 47.7 [39.4, 56.0] n=149 |
| Compatible grade coverage | 30.3 [20.2, 41.9] n=76 | 50.0 [38.3, 61.7] n=76 | 71.1 [59.5, 80.9] n=76 |
| Reference entity availability | 89.4 [83.4, 93.8] n=151 | 89.4 [83.4, 93.8] n=151 | 89.4 [83.4, 93.8] n=151 |
| Metric | MedGemma | PLIP-guided Qwen | PRISM2 |
|---|---|---|---|
| Broad diagnostic agreement | 74.0 [65.3, 81.5] n=123 | 91.1 [84.6, 95.5] n=123 | 100.0 [97.0, 100.0] n=123 |
| Malignancy agreement | 68.1 [59.8, 75.6] n=144 | 77.8 [70.1, 84.3] n=144 | 84.7 [77.8, 90.2] n=144 |
| Invasion agreement | 55.9 [47.4, 64.2] n=143 | 48.3 [39.8, 56.8] n=143 | NA |
| Necrosis agreement | 18.1 [12.3, 25.3] n=149 | 0.7 [0.0, 3.7] n=149 | NA |
| Compatible grade coverage | 44.7 [33.3, 56.6] n=76 | 92.1 [83.6, 97.0] n=76 | NA |
| Reference entity availability | 89.4 [83.4, 93.8] n=151 | 89.4 [83.4, 93.8] n=151 | 89.4 [83.4, 93.8] n=151 |
| Metric | PathXAI-Report (core) | PathXAI-Report |
|---|---|---|
| Broad diagnostic agreement | 100.0 [97.0, 100.0] n=123 | 100.0 [97.0, 100.0] n=123 |
| Malignancy agreement | 84.7 [77.8, 90.2] n=144 | 84.7 [77.8, 90.2] n=144 |
| Invasion agreement | 74.1 [66.1, 81.1] n=143 | 79.7 [72.2, 86.0] n=143 |
| Necrosis agreement | 55.7 [47.3, 63.8] n=149 | 71.8 [63.9, 78.9] n=149 |
| Compatible grade coverage | 35.5 [24.9, 47.3] n=76 | 75.0 [63.7, 84.2] n=76 |
| Reference entity availability | 89.4 [83.4, 93.8] n=151 | 89.4 [83.4, 93.8] n=151 |
| Metric | Qwen+H&E | DeepSeek+H&E | GLM+H&E |
|---|---|---|---|
| Broad diagnostic agreement | 98.4 [94.2, 99.8] n=123 | 95.1 [89.7, 98.2] n=123 | 84.6 [76.9, 90.4] n=123 |
| Malignancy agreement | 83.1 [76.1, 88.8] n=148 | 80.4 [73.1, 86.5] n=148 | 73.6 [65.8, 80.5] n=148 |
| Invasion agreement | 67.8 [59.6, 75.3] n=146 | 76.7 [69.0, 83.3] n=146 | 67.8 [59.6, 75.3] n=146 |
| Necrosis agreement | 70.4 [61.9, 77.9] n=135 | 48.9 [40.2, 57.6] n=135 | 50.4 [41.6, 59.1] n=135 |
| Compatible grade coverage | 31.6 [21.4, 43.3] n=76 | 50.0 [38.3, 61.7] n=76 | 71.1 [59.5, 80.9] n=76 |
| Reference entity availability | 90.1 [84.1, 94.3] n=151 | 90.1 [84.1, 94.3] n=151 | 90.1 [84.1, 94.3] n=151 |
| Metric | MedGemma | PLIP-guided Qwen | PRISM2 |
|---|---|---|---|
| Broad diagnostic agreement | 74.0 [65.3, 81.5] n=123 | 91.1 [84.6, 95.5] n=123 | 100.0 [97.0, 100.0] n=123 |
| Malignancy agreement | 66.2 [58.0, 73.8] n=148 | 76.4 [68.7, 82.9] n=148 | 83.1 [76.1, 88.8] n=148 |
| Invasion agreement | 54.8 [46.4, 63.0] n=146 | 47.9 [39.6, 56.4] n=146 | NA |
| Necrosis agreement | 19.3 [13.0, 26.9] n=135 | 0.7 [0.0, 4.1] n=135 | NA |
| Compatible grade coverage | 44.7 [33.3, 56.6] n=76 | 92.1 [83.6, 97.0] n=76 | NA |
| Reference entity availability | 90.1 [84.1, 94.3] n=151 | 90.1 [84.1, 94.3] n=151 | 90.1 [84.1, 94.3] n=151 |
| Metric | PathXAI-Report (core) | PathXAI-Report |
|---|---|---|
| Broad diagnostic agreement | 100.0 [97.0, 100.0] n=123 | 100.0 [97.0, 100.0] n=123 |
| Malignancy agreement | 83.8 [76.8, 89.3] n=148 | 83.8 [76.8, 89.3] n=148 |
| Invasion agreement | 74.0 [66.1, 80.9] n=146 | 79.5 [72.0, 85.7] n=146 |
| Necrosis agreement | 57.0 [48.2, 65.5] n=135 | 74.1 [65.8, 81.2] n=135 |
| Compatible grade coverage | 36.8 [26.1, 48.7] n=76 | 76.3 [65.2, 85.3] n=76 |
| Reference entity availability | 90.1 [84.1, 94.3] n=151 | 90.1 [84.1, 94.3] n=151 |
| Reference | Method | System | Balanced grade accuracy [95% CI] | n | Case-average exact grade agreement | Grade macro-F1 |
|---|---|---|---|---|---|---|
| P1 | Qwen+H&E | Histological (1–3) | 13.0 [3.9, 57.7] | 76 | 17.1 | 18.3 |
| P1 | DeepSeek+H&E | Histological (1–3) | 19.3 [8.0, 63.8] | 76 | 22.4 | 20.5 |
| P1 | GLM+H&E | Histological (1–3) | 23.1 [10.9, 68.5] | 76 | 34.2 | 28.8 |
| P1 | MedGemma | Histological (1–3) | 18.1 [8.6, 60.4] | 76 | 15.8 | 14.8 |
| P1 | PLIP-guided Qwen | Histological (1–3) | 31.7 [23.9, 66.5] | 76 | 26.3 | 14.7 |
| P1 | PathXAI-Report (core) | Histological (1–3) | 17.1 [6.5, 62.3] | 76 | 23.7 | 23.9 |
| Reference | Method | System | Ordinal grade agreement [95% CI] |
|---|---|---|---|
| P1 | Qwen+H&E | Histological (1–3) | 23.7 [13.8, 36.2] n=76 |
| P1 | DeepSeek+H&E | Histological (1–3) | 35.5 [24.1, 48.5] n=76 |
| P1 | GLM+H&E | Histological (1–3) | 51.3 [38.8, 63.7] n=76 |
| P1 | MedGemma | Histological (1–3) | 30.3 [19.7, 42.8] n=76 |
| P1 | PLIP-guided Qwen | Histological (1–3) | 57.9 [47.3, 67.6] n=76 |
| P1 | PathXAI-Report (core) | Histological (1–3) | 29.6 [18.6, 42.7] n=76 |
| Comparator | Metric | n | Difference [95% CI] |
|---|---|---|---|
| Qwen+H&E | Exact diagnosis match rate | 126 | 13.5 [5.6, 21.4] |
| DeepSeek+H&E | Exact diagnosis match rate | 126 | 9.5 [0.8, 18.3] |
| GLM+H&E | Exact diagnosis match rate | 126 | 24.6 [13.9, 35.7] |
| MedGemma | Exact diagnosis match rate | 126 | 66.7 [57.9, 74.6] |
| PLIP-guided Qwen | Exact diagnosis match rate | 126 | 35.7 [27.0, 44.5] |
| PRISM2 | Exact diagnosis match rate | 126 | 14.3 [4.0, 24.6] |
| Metric | Reference | HistGen | HistoGPT | PathXAI-Report (core) | PathXAI-Report |
|---|---|---|---|---|---|
| Exact diagnosis match rate | P1 | 5.2 [2.1, 10.4] n=135 | 0.0 [0.0, 2.7] n=135 | 57.8 [49.0, 66.2] n=135 | 67.4 [58.8, 75.2] n=135 |
| Finding F1 | P1 | 21.2 [16.3, 26.1] n=150 | 9.2 [5.6, 13.2] n=150 | 67.3 [62.0, 72.7] n=150 | 76.6 [71.2, 81.6] n=150 |
| Finding coverage (A/D) | P1 | 20.7 [16.7, 24.7] n=150 | 14.7 [10.7, 18.7] n=150 | 84.3 [80.0, 88.3] n=150 | 92.7 [89.3, 95.7] n=150 |
| Finding recall (C/D) | P1 | 16.0 [12.3, 19.7] n=150 | 7.3 [4.3, 10.7] n=150 | 64.0 [58.3, 69.3] n=150 | 75.0 [69.7, 80.3] n=150 |
| Finding contradiction rate (E/A) | P1 | 22.1 [12.3, 32.8] n=61 | 51.2 [36.2, 66.2] n=40 | 24.5 [18.7, 30.3] n=147 | 18.7 [13.9, 24.1] n=147 |
| Balanced grade accuracy (histological 1–3) | P1 | 14.9 [6.1, 57.8] n=76 | 0.0 [0.0, 40.0] n=76 | 17.1 [6.5, 62.3] n=76 | 27.4 [ 15.8 , 72.4 ] n=76 |
| Metric | HistGen | HistoGPT | PathXAI-Report (core) | PathXAI-Report |
|---|---|---|---|---|
| Exact grade agreement (sensitivity) | 13.2 [6.5, 22.9] n=76 | 0.0 [0.0, 4.7] n=76 | 23.7 [14.7, 34.8] n=76 | 47.4 [ 35.8 , 59.2 ] n=76 |
| Ordinal grade agreement (sensitivity) | 23.0 [13.5, 35.3] n=76 | 0.0 [0.0, 5.6] n=76 | 29.6 [ 18.6 , 42.7 ] n=76 | 61.2 [ 48.2 , 73.0 ] n=76 |
| Metric | Reference | HistGen | HistoGPT | PathXAI-Report (core) | PathXAI-Report |
|---|---|---|---|---|---|
| Exact diagnosis match rate | P2 | 5.1 [2.1, 10.3] n=136 | 0.0 [0.0, 2.7] n=136 | 55.9 [47.1, 64.4] n=136 | 67.6 [59.1, 75.4] n=136 |
| Finding F1 | P2 | 21.5 [16.3, 26.6] n=151 | 9.5 [6.0, 13.6] n=151 | 70.9 [65.7, 76.0] n=151 | 78.9 [73.7, 83.6] n=151 |
| Finding coverage (A/D) | P2 | 22.0 [17.8, 26.4] n=151 | 14.8 [10.9, 18.8] n=151 | 84.1 [79.7, 88.3] n=151 | 91.5 [88.1, 94.6] n=151 |
| Finding recall (C/D) | P2 | 16.9 [12.9, 21.2] n=151 | 7.6 [4.6, 10.9] n=151 | 67.5 [62.5, 72.7] n=151 | 77.3 [72.1, 82.3] n=151 |
| Finding contradiction rate (E/A) | P2 | 26.2 [15.9, 36.5] n=63 | 50.0 [35.0, 65.0] n=40 | 19.0 [14.2, 24.0] n=145 | 16.2 [11.8, 21.3] n=148 |
| Balanced grade accuracy (histological 1–3) | P2 | 10.8 [2.2, 52.5] n=76 | 0.0 [0.0, 39.4] n=76 | 24.1 [ 9.6 , 65.9 ] n=76 | 42.2 [ 25.1 , 82.7 ] n=76 |
| Metric | HistGen | HistoGPT | PathXAI-Report (core) | PathXAI-Report |
|---|---|---|---|---|
| Exact grade agreement (sensitivity) | 6.6 [2.2, 14.7] n=76 | 0.0 [0.0, 4.7] n=76 | 23.7 [ 14.7 , 34.8 ] n=76 | 61.8 [ 50.0 , 72.8 ] n=76 |
| Ordinal grade agreement (sensitivity) | 19.7 [11.5, 31.1] n=76 | 0.0 [0.0, 5.6] n=76 | 30.3 [ 19.2 , 43.4 ] n=76 | 69.1 [ 56.0 , 80.2 ] n=76 |
| Metric | Reference | HistGen | HistoGPT | PathXAI-Report (core) | PathXAI-Report |
|---|---|---|---|---|---|
| Exact diagnosis match rate | P1 | 5.1 [2.1, 10.3] n=136 | 0.0 [0.0, 2.7] n=136 | 57.4 [48.6, 65.8] n=136 | 66.9 [58.3, 74.7] n=136 |
| Finding F1 | P1 | 21.1 [16.2, 25.9] n=151 | 9.2 [5.6, 13.1] n=151 | 66.9 [61.4, 72.2] n=151 | 76.0 [70.8, 81.0] n=151 |
| Finding coverage (A/D) | P1 | 20.9 [16.9, 25.2] n=151 | 14.6 [10.6, 18.9] n=151 | 84.4 [80.5, 88.4] n=151 | 92.7 [89.4, 95.7] n=151 |
| Finding recall (C/D) | P1 | 15.9 [12.3, 19.5] n=151 | 7.3 [4.3, 10.6] n=151 | 63.6 [57.9, 68.9] n=151 | 74.5 [68.9, 79.5] n=151 |
| Finding contradiction rate (E/A) | P1 | 23.4 [12.9, 33.9] n=62 | 51.2 [36.2, 66.2] n=40 | 25.0 [19.6, 30.7] n=148 | 19.3 [14.2, 24.7] n=148 |
| Balanced grade accuracy (histological 1–3) | P1 | 14.9 [6.1, 57.8] n=76 | 0.0 [0.0, 40.0] n=76 | 17.1 [6.5, 62.3] n=76 | 27.4 [ 15.8 , 72.4 ] n=76 |
| Metric | HistGen | HistoGPT | PathXAI-Report (core) | PathXAI-Report |
|---|---|---|---|---|
| Exact grade agreement (sensitivity) | 13.2 [6.5, 22.9] n=76 | 0.0 [0.0, 4.7] n=76 | 23.7 [14.7, 34.8] n=76 | 47.4 [ 35.8 , 59.2 ] n=76 |
| Ordinal grade agreement (sensitivity) | 23.0 [13.5, 35.3] n=76 | 0.0 [0.0, 5.6] n=76 | 29.6 [ 18.6 , 42.7 ] n=76 | 61.2 [ 48.2 , 73.0 ] n=76 |
| Metric | Reference | HistGen | HistoGPT | PathXAI-Report (core) | PathXAI-Report |
|---|---|---|---|---|---|
| Exact diagnosis match rate | P2 | 5.1 [2.1, 10.2] n=137 | 0.0 [0.0, 2.7] n=137 | 55.5 [46.7, 64.0] n=137 | 67.2 [58.6, 74.9] n=137 |
| Finding F1 | P2 | 21.4 [16.3, 26.8] n=152 | 9.4 [5.8, 13.5] n=152 | 70.5 [65.2, 75.6] n=152 | 78.4 [73.2, 83.2] n=152 |
| Finding coverage (A/D) | P2 | 22.1 [17.9, 26.8] n=152 | 14.7 [10.5, 18.9] n=152 | 84.2 [79.8, 88.4] n=152 | 91.6 [88.2, 94.7] n=152 |
| Finding recall (C/D) | P2 | 16.8 [12.8, 21.1] n=152 | 7.6 [4.6, 10.9] n=152 | 67.1 [61.8, 72.3] n=152 | 76.8 [71.4, 81.9] n=152 |
| Finding contradiction rate (E/A) | P2 | 27.3 [16.4, 38.3] n=64 | 50.0 [35.0, 65.0] n=40 | 19.5 [14.6, 24.4] n=146 | 16.8 [12.1, 21.5] n=149 |
| Balanced grade accuracy (histological 1–3) | P2 | 10.8 [2.2, 52.5] n=76 | 0.0 [0.0, 39.4] n=76 | 24.1 [ 9.6 , 65.9 ] n=76 | 42.2 [ 25.1 , 82.7 ] n=76 |
| Metric | HistGen | HistoGPT | PathXAI-Report (core) | PathXAI-Report |
|---|---|---|---|---|
| Exact grade agreement (sensitivity) | 6.6 [2.2, 14.7] n=76 | 0.0 [0.0, 4.7] n=76 | 23.7 [ 14.7 , 34.8 ] n=76 | 61.8 [ 50.0 , 72.8 ] n=76 |
| Ordinal grade agreement (sensitivity) | 19.7 [11.5, 31.1] n=76 | 0.0 [0.0, 5.6] n=76 | 30.3 [ 19.2 , 43.4 ] n=76 | 69.1 [ 56.0 , 80.2 ] n=76 |
| Outcome | n | MedGemma | Qwen+H&E | PLIP-guided Qwen | PathXAI-Report (core) | PathXAI-Report |
|---|---|---|---|---|---|---|
| Mean rank | 132 | 2.76 [2.63, 2.88] | 1.63 [1.53, 1.74] | 3.29 [3.16, 3.41] | 1.60 [1.50, 1.70] | 1.64 [1.57, 1.72] |
| No correction | 264 | 41.29 | 45.08 | 39.39 | 45.08 | 45.08 |
| Minor correction | 264 | 46.59 | 28.03 | 10.23 | 29.17 | 38.26 |
| Major correction | 264 | 6.06 | 21.21 | 3.41 | 20.45 | 12.88 |
| Unusable | 264 | 6.06 | 5.68 | 46.97 | 5.30 | 3.79 |
| No / minor correction | 264 | 87.88 [83.33, 92.05] | 73.11 [67.80, 78.03] | 49.62 [45.45, 53.79] | 74.24 [69.31, 79.17] | 83.33 [78.79, 87.88] |
| Method | Reader 1 | Reader 2 | Exact agreement (%) | Weighted [95% CI] |
|---|---|---|---|---|
| MedGemma | 1 / 114 / 7 / 10 | 108 / 9 / 9 / 6 | 9.8 | 0.077 [-0.007, 0.167] |
| Qwen+H&E | 1 / 69 / 52 / 10 | 118 / 5 / 4 / 5 | 6.8 | 0.057 [0.002, 0.119] |
| PLIP-guided Qwen | 1 / 15 / 2 / 114 | 103 / 12 / 7 / 10 | 10.6 | 0.017 [0.001, 0.036] |
| PathXAI-Report (core) | 1 / 73 / 50 / 8 | 118 / 4 / 4 / 6 | 8.3 | 0.077 [0.015, 0.145] |
| PathXAI-Report | 0 / 97 / 30 / 5 | 119 / 4 / 4 / 5 | 5.3 | 0.065 [0.002, 0.133] |
| Method | No/minor correction | Cancer family match | Exact diagnosis match | Family match among no/minor | Exact match among no/minor |
|---|---|---|---|---|---|
| MedGemma | 232/264 (87.9%) | 14/132 (10.6%) | 4/99 (4.0%) | 27/232 (11.6%) | 7/168 (4.2%) |
| Qwen+H&E | 193/264 (73.1%) | 102/132 (77.3%) | 65/99 (65.7%) | 150/193 (77.7%) | 98/146 (67.1%) |
| PLIP-guided Qwen | 131/264 (49.6%) | 44/132 (33.3%) | 24/99 (24.2%) | 51/131 (38.9%) | 30/100 (30.0%) |
| PathXAI-Report (core) | 196/264 (74.2%) | 104/132 (78.8%) | 68/99 (68.7%) | 157/196 (80.1%) | 94/142 (66.2%) |
| PathXAI-Report | 220/264 (83.3%) | 96/132 (72.7%) | 60/99 (60.6%) | 164/220 (74.5%) | 103/162 (63.6%) |
| Metric | Qwen+H&E | H&E+tissue statistics | H&E+segmentation overlays | PathXAI-Report |
|---|---|---|---|---|
| Exact diagnosis match rate | 58.9 [54.9, 62.9] n=604 | 61.1 [57.1, 65.0] n=604 | 56.1 [52.1, 60.1] n=604 | 62.3 [58.3, 66.1] n=604 |
| Diagnostic composite (cohort target) | 84.8 [83.5, 86.0] n=1200 | 84.9 [83.6, 86.1] n=1200 | 82.7 [81.4, 84.0] n=1200 | 86.9 [85.8, 88.0] n=1200 |
| Cancer family match rate | 68.1 [65.4, 70.7] n=1200 | 69.2 [66.5, 71.8] n=1200 | 64.6 [61.8, 67.3] n=1200 | 72.5 [69.9, 75.0] n=1200 |
| Broad diagnostic agreement | 98.9 [98.2, 99.4] n=1200 | 97.6 [96.5, 98.4] n=1200 | 97.8 [96.8, 98.6] n=1200 | 99.2 [98.5, 99.6] n=1200 |
| Malignancy agreement | 90.8 [89.0, 92.3] n=1200 | 91.8 [90.1, 93.3] n=1200 | 89.8 [88.0, 91.5] n=1200 | 91.9 [90.2, 93.4] n=1200 |
| Invasion agreement | 71.6 [67.9, 75.1] n=641 | 69.3 [65.5, 72.8] n=641 | 69.1 [65.4, 72.7] n=641 | 78.6 [75.2, 81.7] n=641 |
| Contrast | Metric | n | Difference [95% CI] |
|---|---|---|---|
| H&E+tissue statistics minus Qwen+H&E | Diagnostic composite (cohort target) | 1200 | 0.1 [-0.9, 1.1] |
| H&E+tissue statistics minus Qwen+H&E | Cancer family match rate | 1200 | 1.1 [-1.0, 3.2] |
| H&E+tissue statistics minus Qwen+H&E | Exact diagnosis match rate | 604 | 2.2 [-0.5, 4.8] |
| H&E+tissue statistics minus Qwen+H&E | Broad diagnostic agreement | 1200 | -1.3 [-2.2, -0.5] |
| H&E+tissue statistics minus Qwen+H&E | Finding F1 | 997 | 0.0 [-2.1, 2.2] |
| H&E+tissue statistics minus Qwen+H&E | Finding coverage (A/D) | 997 | -2.6 [-4.6, -0.5] |
| Metric | All 456 cases: difference (n) | 443 flagged-available cases: difference (n); 95% CI |
|---|---|---|
| H&E+tissue statistics minus H&E only | ||
| Cancer family match rate | +4.17 (456) | +4.06 (443); [0.90, 7.22] |
| Exact diagnosis match rate | +4.55 (352) | +4.71 (340); [0.59, 8.53] |
| Finding F1 | -2.80 (408) | -2.95 (397); [-6.42, 0.41] |
| Finding recall | -2.72 (408) | -2.87 (397); [-6.00, 0.21] |
| Finding contradiction rate | -2.03 (251) | -2.10 (242); [-5.41, 1.21] |
| Metric / grading system | n | Difference [95% CI] |
|---|---|---|
| Panel 4.3. Report-workflow comparisons : Tissue statistics at writing only minus positive tissue statistics | ||
| Diagnostic composite (cohort target) | 42 | -3.9 [-9.8, 1.0] |
| Cancer family match rate | 42 | -7.1 [-16.7, 2.4] |
| Exact diagnosis match rate | 33 | -3.0 [-12.1, 6.1] |
| Finding F1 | 34 | -4.3 [-16.4, 8.1] |
| Balanced grade accuracy; Fuhrman (1–4) | 1 | 0.0 |
| Comparison | Cases | Original weights | Equal weights |
|---|---|---|---|
| Public evaluation cohort, reference-diagnosis composite | |||
| PathXAI-Report vs Qwen+H&E | 1200 | +1.17 [0.31, 2.08] | +1.19 [0.40, 2.03] |
| PathXAI-Report vs PathXAI-Report (core) | 1200 | +1.61 [0.83, 2.44] | +1.35 [0.65, 2.08] |
| Public evaluation cohort, cohort-target composite | |||
| PathXAI-Report vs Qwen+H&E | 1200 | +2.13 [0.97, 3.15] | +1.94 [1.00, 2.81] |
| PathXAI-Report vs PathXAI-Report (core) | 1200 | +2.24 [1.26, 3.22] | +1.89 [1.08, 2.72] |
| Method | Primary (%) | With summary (%) | Change, pp [95% CI] | Added/n | Corrected/ lost |
|---|---|---|---|---|---|
| PathXAI-Report pipeline development subset (n= 352 ) | |||||
| Qwen+H&E | {\color[rgb]{0,0,0}61.65} | {\color[rgb]{0,0,0}59.94} | {\color[rgb]{0,0,0}-1.70} [ , {\color[rgb]{0,0,0}-0.57} ] | 352 / 352 | 0 / 6 |
| DeepSeek+H&E | {\color[rgb]{0,0,0}62.50} | {\color[rgb]{0,0,0}61.08} | {\color[rgb]{0,0,0}-1.42} [ {\color[rgb]{0,0,0}-2.84} , {\color[rgb]{0,0,0}-0.28} ] | 352 / 352 | 0 / 5 |
| GLM+H&E | {\color[rgb]{0,0,0}46.59} | {\color[rgb]{0,0,0}46.02} | {\color[rgb]{0,0,0}-0.57} [ {\color[rgb]{0,0,0}-3.41} , {\color[rgb]{0,0,0}2.27} ] | 352 / 352 | 11 / 13 |
| MedGemma | {\color[rgb]{0,0,0}3.69} | {\color[rgb]{0,0,0}7.39} | {\color[rgb]{0,0,0}3.69} [ {\color[rgb]{0,0,0}1.99} , {\color[rgb]{0,0,0}5.68} ] | 349 / 352 | 13 / 0 |
| PLIP-guided Qwen | {\color[rgb]{0,0,0}23.86} | {\color[rgb]{0,0,0}23.86} | [ , ] | 0/ 352 | 0/0 |
| Comparison | n | Primary difference | With summary, difference [95% CI] |
|---|---|---|---|
| PathXAI-Report pipeline development subset | |||
| PathXAI-Report vs PathXAI-Report (core) | 352 | {\color[rgb]{0,0,0}-0.57} [ {\color[rgb]{0,0,0}-4.55} , {\color[rgb]{0,0,0}3.41} ] | |
| PathXAI-Report vs Qwen+H&E | 352 | {\color[rgb]{0,0,0}4.55} | {\color[rgb]{0,0,0}4.83} [ {\color[rgb]{0,0,0}0.00} , {\color[rgb]{0,0,0}9.38} ] |
| Public evaluation cohort | |||
| PathXAI-Report vs PathXAI-Report (core) | 604 | {\color[rgb]{0,0,0}4.14} | {\color[rgb]{0,0,0}3.31} [ {\color[rgb]{0,0,0}0.66} , {\color[rgb]{0,0,0}6.13} ] |
| PathXAI-Report vs Qwen+H&E | 604 | {\color[rgb]{0,0,0}3.31} | {\color[rgb]{0,0,0}2.48} [ {\color[rgb]{0,0,0}-0.83} , {\color[rgb]{0,0,0}5.79} ] |
| Method | n | Primary (%) | With summary (%) | Change, pp [95% CI] |
|---|---|---|---|---|
| Public pooled cohort | ||||
| Qwen+H&E | 1656 | {\color[rgb]{0,0,0}69.02} | {\color[rgb]{0,0,0}68.78} | {\color[rgb]{0,0,0}-0.24} [ {\color[rgb]{0,0,0}-0.79} , {\color[rgb]{0,0,0}0.30} ] |
| DeepSeek+H&E | 1656 | {\color[rgb]{0,0,0}71.56} | {\color[rgb]{0,0,0}71.14} | {\color[rgb]{0,0,0}-0.42} [ {\color[rgb]{0,0,0}-0.91} , {\color[rgb]{0,0,0}0.00} ] |
| GLM+H&E | 1656 | {\color[rgb]{0,0,0}60.21} | {\color[rgb]{0,0,0}57.37} | {\color[rgb]{0,0,0}-2.84} [ {\color[rgb]{0,0,0}-4.23} , {\color[rgb]{0,0,0}-1.45} ] |
| MedGemma | 1656 | {\color[rgb]{0,0,0}6.94} | {\color[rgb]{0,0,0}12.20} | {\color[rgb]{0,0,0}5.25} [ {\color[rgb]{0,0,0}4.17} , {\color[rgb]{0,0,0}6.34} ] |
| PLIP-guided Qwen | 1656 | {\color[rgb]{0,0,0}41.30} | {\color[rgb]{0,0,0}41.30} | [ , ] |
| Comparison | n | Primary difference | With summary, difference [95% CI] |
|---|---|---|---|
| Public pooled cohort | |||
| PathXAI-Report vs DeepSeek+H&E | 1656 | {\color[rgb]{0,0,0}2.11} | {\color[rgb]{0,0,0}0.42} [ {\color[rgb]{0,0,0}-1.75} , {\color[rgb]{0,0,0}2.66} ] |
| PathXAI-Report vs PathXAI-Report (core) | 1656 | {\color[rgb]{0,0,0}3.99} | {\color[rgb]{0,0,0}1.75} [ {\color[rgb]{0,0,0}-0.12} , {\color[rgb]{0,0,0}3.62} ] |
| PathXAI-Report vs Qwen+H&E | 1656 | {\color[rgb]{0,0,0}4.65} | {\color[rgb]{0,0,0}2.78} [ {\color[rgb]{0,0,0}0.66} , {\color[rgb]{0,0,0}4.89} ] |
| PathXAI-Report pipeline development subset | |||
| PathXAI-Report vs Qwen+H&E | 456 | {\color[rgb]{0,0,0}5.26} | {\color[rgb]{0,0,0}3.29} [ {\color[rgb]{0,0,0}-0.66} , {\color[rgb]{0,0,0}7.46} ] |
| Comparator | Cancer family match rate | Exact diagnosis match rate | Finding F1 | Finding recall | Finding contradiction rate |
|---|---|---|---|---|---|
| Public evaluation cohort (1,200 cases) | |||||
| Qwen+H&E | +4.42 (0.024) | +3.31 (1.000) | +5.75 ( ) | +5.07 ( ) | -5.52 ( ) |
| DeepSeek+H&E | +0.58 (1.000) | +1.99 (1.000) | +6.67 ( ) | +5.49 ( ) | -9.19 ( ) |
| GLM+H&E | +10.67 ( ) | +6.29 (0.200) | +4.60 (0.030) | +3.50 (0.232) | -10.01 ( ) |
| MedGemma | +65.83 ( ) | +56.13 ( ) | +11.33 ( ) | +9.86 ( ) | -1.08 (1.000) |
| PLIP-guided Qwen | +26.83 ( ) | +23.01 ( ) | +26.88 ( ) | +23.87 ( ) | +4.75 (0.035) |
| Method | Public evaluation | Dedicated-generator public comparison | Private cohort |
|---|---|---|---|
| Qwen+H&E | 1200 / 1200 | — | 152 / 152 |
| DeepSeek+H&E | 1200 / 1200 | — | 152 / 152 |
| GLM+H&E | 1200 / 1200 | — | 151 / 152 |
| MedGemma | 1200 / 1200 | — | 152 / 152 |
| PLIP-guided Qwen | 1200 / 1200 | — | 152 / 152 |
| PRISM2 | 1200 / 1200 | — | 152 / 152 |
| Comparator | Cancer family match rate | Exact diagnosis match rate | Finding F1 | Finding recall |
|---|---|---|---|---|
| Public dedicated-generator comparison: 215 shared / 217 eligible cases | ||||
| HistGen | +38.60 / +38.25 | +20.55 / +20.55 | +26.06 / +26.14 | +22.40 / +22.52 |
| HistoGPT | +72.56 / +71.89 | +61.64 / +61.64 | +31.28 / +31.31 | +27.18 / +27.25 |
| PathXAI-Report (core) | +4.19 / +3.23 | +4.11 / +4.11 | +0.07 / +0.17 | +0.19 / +0.29 |
| Private reference set P1: 151 shared / 152 eligible cases | ||||
| Qwen+H&E | +4.13 / +4.13 | +13.33 / +13.24 | +5.00 / +4.64 | +7.00 / +6.62 |
| Comparator | Cancer family match rate | Exact diagnosis match rate | Finding F1 | Finding recall | Finding contradiction rate |
|---|---|---|---|---|---|
| Public evaluation cohort (1,200 cases) | |||||
| Qwen+H&E | +4.42 (1200) [2.08, 6.83] | +3.31 (604) [0.33, 6.29] | +5.75 (997) [3.66, 7.87] | +5.07 (997) [3.10, 7.07] | -5.52 (664) [-8.01, -3.15] |
| DeepSeek+H&E | +0.58 (1200) [-1.92, 3.00] | +1.99 (604) [-1.66, 5.79] | +6.67 (997) [4.17, 9.15] | +5.49 (997) [3.19, 7.81] | -9.19 (634) [-12.10, -6.31] |
| GLM+H&E | +10.67 (1200) [7.83, 13.50] | +6.29 (604) [2.15, 10.10] | +4.60 (997) [2.07, 7.17] | +3.50 (997) [0.91, 5.95] | -10.01 (644) [-13.02, -6.99] |
| MedGemma | +65.83 (1200) [63.50, 68.08] | +56.13 (604) [52.65, 59.27] | +11.33 (997) [8.89, 13.78] | +9.86 (997) [7.58, 12.17] | -1.08 (546) [-3.77, 1.57] |
| PLIP-guided Qwen | +26.83 (1200) [24.67, 29.00] | +23.01 (604) [20.03, 26.16] | +26.88 (997) [24.68, 29.16] | +23.87 (997) [21.72, 26.01] | +4.75 (327) [2.09, 7.41] |
| Method | Internal 5 test sets | External 6 tasks | Combined 11 test sets |
|---|---|---|---|
| HoVer-Net FT, aligned | 62.02 [58.29, 65.87] | 33.95 [32.60, 35.27] | 46.71 [44.85, 48.59] |
| nnU-Net 2D, aligned | 64.43 [60.19, 68.64] | 39.21 [37.80, 40.51] | 50.67 [48.61, 52.70] |
| CONCHSeg, aligned | 14.22 [12.72, 15.70] | 26.44 [24.81, 27.77] | 20.88 [19.76, 21.86] |
| MedSAM3 Zero-shot | 31.45 [29.50, 33.39] | 56.34 [54.08, 59.05] | 45.02 [43.51, 46.74] |
| SAM2 FT | 63.38 [58.52, 68.42] | 71.63 [70.15, 73.24] | 67.88 [65.52, 70.34] |
| MedSAM3 FT | 65.67 [60.65, 70.84] | 72.28 [70.42, 74.28] | 69.28 [66.78, 71.88] |
| Dataset | PathSegmentor | VISTA-PATH | PathXAI-Seg (SAM2) | PathXAI-Seg (MedSAM3) |
|---|---|---|---|---|
| ACDC-LungHP | 82.44 [78.20, 85.48] | 84.45 [80.83, 87.10] | 85.06 [81.58, 87.64] | 84.73 [81.15, 87.39] |
| PAIP2019 | 70.14 [67.34, 73.09] | 71.61 [68.43, 74.76] | 76.07 [73.74, 78.51] | 75.67 [72.99, 78.38] |
| METU-CCTGS | 21.79 [15.29, 28.55] | 69.41 [62.33, 76.17] | 72.21 [66.42, 77.89] | 69.79 [62.86, 76.54] |
| Shared-set macro | 58.12 [55.34, 60.80] | 75.16 [72.35, 77.82] | 77.78 [75.43, 80.03] | 76.73 [74.01, 79.34] |
| Test / weighting | SAM2 gain | MedSAM3 gain |
|---|---|---|
| internal; tasks | 1.66 [0.14, 3.04] | 2.59 [0.16, 5.20] |
| internal; organs | 3.04 [0.99, 4.71] | 3.56 [-0.11, 7.63] |
| external; tasks | 0.15 [-0.28, 0.66] | 0.83 [0.44, 1.32] |
| external; organs | 0.33 [0.01, 0.70] | 0.79 [0.36, 1.37] |
| Test / organ | SAM2 FT | PathXAI-Seg (SAM2) | MedSAM3 FT | PathXAI-Seg (MedSAM3) |
|---|---|---|---|---|
| Internal / Breast | 60.77 [51.23, 70.42] | 64.42 [54.02, 74.88] | 70.43 [63.38, 77.58] | 72.14 [66.60, 77.72] |
| Internal / Colorectal | 77.41 [74.30, 80.28] | 72.93 [69.91, 75.85] | 71.28 [67.67, 74.86] | 71.84 [67.89, 75.72] |
| Internal / Lung | 40.56 [30.26, 52.55] | 50.49 [42.21, 59.47] | 44.92 [26.57, 63.68] | 53.36 [40.61, 66.28] |
| External / Colorectal | 71.91 [66.51, 77.33] | 72.21 [66.59, 77.91] | 67.53 [59.96, 75.11] | 69.79 [63.20, 76.61] |
| External / Liver | 75.16 [72.74, 77.69] | 76.07 [73.68, 78.55] | 75.89 [73.36, 78.48] | 75.67 [72.97, 78.39] |
| External / Lung | 84.72 [81.05, 87.50] | 85.06 [81.58, 87.67] | 84.50 [80.85, 87.26] | 84.73 [81.16, 87.41] |
| Class / group | HoVer-Net | nnU-Net | CONCHSeg | MedSAM3 | SAM2 | MedSAM3 | PathXAI-Seg | ||
|---|---|---|---|---|---|---|---|---|---|
| FT | zero-shot | FT | FT | SAM2 | MedSAM3 | ||||
| BCSS (internal) | |||||||||
| Neoplastic | 31 | 56.79 | 72.52 | 50.95 | 0.31 | 77.74 | 81.58 | 79.88 | 82.01 |
| Inflammatory | 26 | 43.89 | 53.99 | 0.00 | 0.00 | 55.07 | 67.10 | 62.86 | 66.78 |
| Connective | 31 | 50.21 | 60.09 | 0.00 | 11.86 | 63.50 | 65.15 | 64.92 | 63.63 |
| Dead / necrotic | 19 | 0.00 | 15.79 | 0.00 | 4.35 | 50.20 | 62.42 | 54.17 | 66.57 |
| Class / group | HoVer-Net | nnU-Net | CONCHSeg | MedSAM3 | SAM2 | MedSAM3 | PathXAI-Seg | ||
|---|---|---|---|---|---|---|---|---|---|
| FT | zero-shot | FT | FT | SAM2 | MedSAM3 | ||||
| BCSS (internal) | |||||||||
| Multiple classes | 31 | 40.27 | 52.46 | 14.67 | 4.47 | 60.74 | 67.42 | 65.57 | 68.51 |
| DigestPath2019 (internal) | |||||||||
| Multiple classes | 50 | 67.07 | 67.71 | 10.99 | 61.61 | 77.03 | 75.73 | 79.28 | 75.98 |
| GlaS (internal) | |||||||||
| Dataset | Class | PathSegmentor | VISTA-PATH | PathXAI-Seg (SAM2) | PathXAI-Seg (MedSAM3) | |
| Class-specific means | ||||||
| ACDC-LungHP | Foreground | 5,000 | 91.29 / 90.02 | 90.17 / 88.88 | 91.29 / 89.79 | 91.62 / 90.31 |
| Background | 751 | 23.50 / 18.71 | 46.32 / 37.59 | 43.53 / 33.49 | 38.88 / 31.40 | |
| PAIP2019 | Foreground | 1,719 | 75.50 / 70.94 | 64.97 / 59.14 | 91.02 / 87.22 | 87.19 / 83.07 |
| Background | 4,946 | 68.27 / 67.75 | 73.92 / 72.13 | 70.88 / 70.19 | 71.66 / 70.81 | |
| METU-CCTGS | Foreground | 16 | 17.49 / 10.30 | 50.24 / 37.12 | 56.08 / 41.33 | 54.36 / 40.27 |
| Scope / metric | Configuration / contrast | n | Estimate | 95% CI |
|---|---|---|---|---|
| Training checkpoints: Internal (5 test sets; n=124 images) | ||||
| Dice | Supervised adaptation | 124 | 65.67 | [60.65, 70.84] |
| Dice | Joint supervision | 124 | 69.29 | [65.61, 73.10] |
| Dice | Focal refinement | 124 | 68.26 | [64.47, 72.09] |
| Training checkpoints: External (6 tasks; n=11089 images/specimens) | ||||
| Dice | Supervised adaptation | 11089 | 72.28 | [70.42, 74.28] |
Medical Report Generation
Medical report generation turns medical images, clinical records, or conversations into structured, clinically useful reports.
See all 90 papers