Semi-supervised learning for automated perineural invasion detection in multi-organ H&E whole slide images

Perineural invasion (PNI) is an important pathological phenotype associated with poor prognosis in multiple malignancies. The primary detection method is visual inspection of whole slide images (WSIs), which is labor-intensive, time-consuming, subjective, and prone to high inter-observer variability. Developing reliable, accurate deep learning models for PNI detection is constrained by the lack of pixel-level annotated WSIs.

We evaluated four backbone architectures and two different approaches to improve PNI detection in a multi-organ dataset of colon, prostate, and pancreatic adenocarcinomas. We report three key findings. First, two pathology-pretrained foundation models, Virchow-2 and UNI, substantially outperformed ImageNet-pretrained CNNs (EfficientNet-B3, ConvNeXt-2), with distinct baseline error profiles reflecting differences in pretraining-data composition. Second, a data curation strategy driven by confidence-based pseudo-labelling (threshold P > .9) with human-in-the-loop review expanded the dataset from 262 to 352 WSIs, yielding a 12.4% relative F1 improvement (0.740 to 0.832) and a 55.5% reduction in false positives per slide; an ablation attributed 70% of the F1 gain to data volume and 41% of the FP reduction to benign-class curation. Third, per-organ analysis revealed that the primary driver of this improvement was not data volume alone but the targeted annotation enrichment of underrepresented morphologies in adjacent-normal and benign tissue, including desmoplastic stroma, crypts, and small blood vessels, that had been a systematic source of false positive predictions across tissue types.

Our implementation is available at https://github.com/AhmadAlkhan/PNI_SSL.