Skin Lesion Classifiers

Skin Lesion Classifiers is a collection of image classification models designed for dermatological applications, specifically for identifying skin conditions from images. The models are available as trained checkpoints and are implemented using PyTorch. This resource includes six distinct classifiers trained on a dataset covering 14 different skin condition classes.

The collection features two custom convolutional neural networks (CNNs) that were trained from scratch, as well as four hybrid models that combine a pretrained vision transformer backbone with Kolmogorov-Arnold Network (KAN) spline layers. Among the hybrid models, two variants introduce an alternating KAN-MLP block, which interleaves KAN and standard multilayer perceptron (MLP) feedforward layers across the transformer blocks, rather than replacing the MLP component entirely. This architectural approach is highlighted as a novel aspect of the collection.

The tool provides access to the trained model checkpoints, with reported validation and test accuracies for at least two of the custom CNN models. Documentation, code, and training notebooks are made available through a linked GitHub repository, supporting reproducibility and further experimentation by users. The models are distributed under the MIT license, allowing for open-source use and modification.

evidence_sufficient": true}