Multi-Resolution Skin Cancer ViT

A dermatopathologist reading a skin biopsy changes objective lenses constantly — low power to find the lesion and judge its architecture, high power to inspect nuclei. A patch classifier trained at one magnification has no such freedom. The basal cell carcinoma nest that fills a 40x field is a smudge of a few dozen pixels at 10x, and a slide from a scanner set to a different objective is, to that model, tissue it has never seen. Laboratories do not standardise on one scanner or one native magnification, so this is a deployment problem rather than an academic one.

This model makes magnification part of the training distribution instead of a fixed assumption. Roughly 1.13 million hematoxylin-and-eosin patches were assembled from two public skin-biopsy collections and rendered at 10x, 20x, 40x and 400x, and a single vision transformer was fine-tuned across all of them, so one checkpoint reads a slide at whatever resolution it arrives in. It was built by Abadh K. Chaurasia and colleagues at the Menzies Institute for Medical Research, University of Tasmania, with the Department of Pathology at Royal Hobart Hospital and Pandani Solutions, and published in Computers in Biology and Medicine in 2025.

Each patch is assigned one of six labels — basal cell carcinoma (BCC), squamous cell carcinoma (SCC), intraepidermal carcinoma (IEC), melanoma, naevi, or non-cancerous — and those calls are rolled up into a whole-slide diagnosis. A reduced variant covering only the non-melanoma classes was then carried unchanged onto a cohort from another country, scanner and magnification.

#Key Features

• Magnification-agnostic patch reading: One set of weights classifies patches sampled at 10x, 20x, 40x and 400x, rather than requiring input resampled to a single fitted scale.
• Six-class dermatopathology taxonomy: The label set spans the three common non-melanoma keratinocyte cancers alongside melanoma, naevi and non-cancerous tissue, so a slide need not be pre-sorted into melanocytic and non-melanocytic.
• Rule-based slide aggregation: A slide is called cancerous when more than 5% of its tissue-bearing patches are, and takes its most frequent cancerous label. The rule is hand-set rather than learned, unlike the attention aggregator in Pan-Cancer TP53 MIL ViT.
• Attention maps per head: Weights from all twelve heads of the final self-attention layer are normalised and overlaid on the patch, giving a visual account of each call.

#Technical Details

The backbone is vit_base_patch14_reg4_dinov2.lvd142m — a ViT-B/14 with four register tokens, roughly 86.6M parameters, taking 518×518 input. Its self-supervised pretraining is inherited rather than performed here: DINOv2 on LVD-142M natural images, not histology. Patches are Macenko stain-normalised, then fastai transfer learning runs the head for ten epochs at a learning rate of 1.2e-3 and unfreezes the whole network for ten more under a one-cycle policy spanning 1e-6 to 1e-4, on an NVIDIA A100. The corpus draws on the Heidelberg dataset (386 cases on an Aperio AT2 at 400x, annotated across 16 tissue categories) and the Non-Melanoma Skin Cancer Segmentation set (290 MyLab Pathology slides — 140 BCC, 60 SCC, 90 IEC — imaged at 10x and resampled to 20x and 40x).

Across six classes the model reaches a quadratic weighted Cohen's κ of 0.859 (95% CI 0.851–0.866) on validation and 0.898 (0.892–0.904) on 161,088 held-out test patches. IEC is the weak class at 64.0% and 70.9% accuracy under heavy class imbalance, and melanoma and naevi training data existed only at 400x, so those three classes were dropped to build the non-melanoma model, which reaches κ 0.919 (0.914–0.924) on validation. That model was then run over the COBRA cohort — 5,147 slides from 4,066 patients at Radboud University Medical Center, digitised on a 3DHistech Pannoramic 1000 at 20x, 2,673 BCC against 2,474 non-BCC — scoring slide-level κ 0.791 (0.774–0.808), accuracy 0.896 and AUROC 0.897, while misclassifying 616 non-cancerous slides as SCC.

#Applications

The intended use is assistive dermatopathology: outlining cancerous regions on a whole slide image before a pathologist reads it, and proposing a subtype with an attention map showing what the call rested on. The authors position the non-melanoma variant for settings with high skin-cancer incidence and limited access to experienced dermatopathologists, where the available scanner is whatever the laboratory already owns.

#Impact

The contribution is evidence that a task-specialised histopathology classifier can survive a change of country, scanner and magnification: prior weakly supervised BCC work by Geijs et al. reported AUC 0.958 on 397 slides, while this checkpoint was tested on twelve times that volume. The limits matter for anyone reading the headline number. Only the non-melanoma variant was externally validated, and only as BCC versus non-BCC; the melanoma, naevi and IEC arm has never been tested on unseen data; training drew on two datasets, which constrains generalisation across populations; and 616 false SCC calls is a specificity problem for triage use. No training code, checkpoint or model licence has been located, and three of the four authors declare co-founding interests in the developing company.