PathJev: Decision-Native Vision–Language Modelling for Histopathology¶
Reference entry only. The full pathology study was not reproduced due to cost (review complete).
Paper information¶
| Field | Details |
|---|---|
| Paper | PathJev: Decision-Native Vision–Language Modelling for Histopathology; reviewed full text |
| Affiliations | First author Cristian Tommasino: University of Naples Federico II |
| First public release | 2026-10-23 10:47:40 UTC (announced 2026-10-26) |
| Author code | CosmoIknosLab/PathJev at 1f5fa8d29220fbdbc31496484170e85fe83d4bea, Apache-2.0, includes implementation, inference, analysis, configs, tests and results. PathJev-P weights are available. Paper CC BY-NC-ND 4.0; MedGemma and data terms are separate. |
| Local key | pathjev; no adapter |
| Local code | No executable local implementation. |
| Local evidence | No local experiment, metrics or GPU receipt. |
What the paper changes¶
PathJev converts expert pathology-patch labels into typed Choice/Noul/Score decisions with candidate sets supplied at runtime. In one forward pass, the shared decision head reads question and candidate marker states, combines candidate, interaction and set-attention terms, and normalizes over the given candidates. The head and two-stage training follow OmniMed-Jev. The pathology contribution is the 40K decision states, typed labeling, patient/slide and overlap audit, unseen-organ tests and generative readout comparisons.
Original paper figure¶
Original figure from the reviewed paper, reused without redrawing. See figure provenance.
Core equations¶
Candidates are input text, not global classifier classes, and the prompt excludes the target. Score uses one-hot grade labels as specified, not a replacement ordinal-distance loss.
Results reported by the authors¶
On the test with no known overlap, balanced accuracy is 0.541 versus 0.492 for a same-backbone generative baseline. Calibration depends on readout: mean ECE is around 0.201 for both length-normalized generative output and PathJev. The paper notes unseen-organ overconfidence and missing patient IDs in some data. No production A/B or local capability result.
Local reproduction and limits¶
No operator, command, metrics or GPU receipt. The author repository is a reference; no MedGemma, patches or whole slides were downloaded, and no generic classifier or existing Jev infrastructure is represented as a PathJev implementation.
Full experiment: not reproduced¶
The study uses MedGemma1.5-4B NF4, 896×896 images, roughly 39.6M LoRA/head parameters and 40K states. Table 7 reports 10.4–10.5 hours per single-GPU PathJev seed versus 16.3–16.5 hours for Gen-SFT, before the added cost of hyperparameter selection, three seeds, multi-dataset inference and whole-slide experiments. No download, trial or retry.
