On the hardest tissue class, a pathology AI was the only one to lead

A head-to-head histology benchmark tested nine deep-learning architectures on one task: classifying microscope images of tissue by type. The training set was 4,307 H&E-stained slides (the standard pink-and-purple histology stain) from male rats across lung, cerebellum, and adipose. The test set was 600 mixed human-and-animal images, a cross-domain check on whether the models could generalize. (PMID 42604434)

On internal cross-validation, every model looked near-perfect, so the internal benchmark could not separate the field. The external cohort is where the architectures diverged.

On easier tissues, adipose and cerebellum, competitive models reached F1 above 95%. On the hardest class (lung, the most structurally diverse), the spread widened. UNI2-h, a foundation model pre-trained on large-scale pathology image collections, led with F1 97.4% and recall 95.0%. General-purpose vision encoders, including DINOv3 ViT variants and ConvNeXt, fell behind. (PMID 42604434)

For compute, MobileNetV3-Small was fastest at 9.4 ms per image on CPU and 12.5 ms on GPU, but speed and accuracy decoupled on the external set.

The paper is narrow: one lab's rat data, 600 external images, no clinical workload. Domain-specific foundation models closed more of the lab-to-real gap than off-the-shelf vision encoders on the hardest tissue class. Internal benchmarks could not rank the field. (PMID 42604434)