Evidence Field: compact, auditable histopathology classification transfers across clinical centers

With only 10,550 parameters per model, Evidence Field couples spatial image measurements with replayable covering programs, surpassing standard CNN and ViT baselines across held-out hospital sites on CAMELYON17-WILDS.

Primary source: arXiv:2610.21115: https://arxiv.org/abs/2610.21115

What’s new: Deep convolutional networks and vision transformers deployed for digital pathology frequently suffer from domain shift across hospital centers due to variations in tissue preparation, staining protocols, and scanner optics. Compounding this brittleness, deep neural networks function as opaque black boxes whose post-hoc saliency maps fail to provide deterministic, verifiable explanations for clinical audit. In Evidence Field for Histopathology: Cross-Centre Evaluation and Replayable Spatial Decisions (arXiv:2610.21115, 15 Oct 2026), researchers Isarar Siddique, Raiyan Siddique, Arshi Niyaz, Jamyang Ugyen Tshomo, and Atish Raza introduce Evidence Field, an ultra-compact classifier architecture that produces natively replayable spatial decision records.

Evidence Field fundamentally re-architects tissue analysis by coupling learned optical image measurements with an adaptive spatial covering program, requiring only 10,550 trainable parameters per ensemble member. The authors evaluated the method across all 455,954 patches in the CAMELYON17-WILDS lymph node metastasis benchmark, using patient-separated source partitions and evaluating generalization on two strictly held-out hospital centers. The source-selected paired-cross-entropy Evidence Field ensemble achieved areas under the ROC curve (AUROC) of 0.9356 and 0.9453, with corresponding classification accuracies of 84.46% and 86.77% on the unseen hospital centers. By contrast, standard residual convolutional networks achieved AUROCs of only 0.7352 and 0.7656, while compact vision transformers reached 0.8502 and 0.7924. Beyond superior domain transfer, Evidence Field outputs an exact computational receipt for every prediction: signed regional and interaction terms, contextual bounding covers, and deterministic spatial programs that were successfully audited and reconstructed across 6,120 class instances under multiple viewing rotations.

Why it matters: Clinical pathology demands high cross-site reliability and strict regulatory auditability, criteria that black-box deep learning models struggle to satisfy simultaneously. Evidence Field proves that parameter-heavy foundation models are not strictly required for robust domain generalization in histopathology: structured spatial programs with four orders of magnitude fewer parameters can outperform standard deep models while providing mathematical traceability for every diagnostic call.

Caveats: The empirical validation was conducted on pre-extracted image patches from the CAMELYON17-WILDS dataset rather than gigapixel whole-slide images with complex macro-architectural context. Performance on rare non-epithelial malignancies or severely understained tissue specimens was not systematically evaluated. Furthermore, generating recursive spatial covering programs introduces structural tiling overhead that may run slower during raw batch inference compared to optimized single-pass convolutional feature extractors.