CANet: A color-aware convolutional neural network with clinically grounded chromatic learning for skin lesion classification from dermoscopic images

Classification of dermoscopic skin lesions relies significantly on color indicators, where red signals inflammation, blue-gray signifies invasive melanoma, and white denotes regression structures. Nonetheless, conventional convolutional neural networks (CNNs) treat RGB channels separately and do not incorporate inductive biases related to the clinical significance of color patterns. Current color-sensitive approaches typically involve basic color space conversions (such as RGB to HSV) or identification of fixed chromatic features (like the blue-white veil) using manually designed rules, failing to capture task-specific color interactions through end-to-end learning. This study presents CANet, a color-aware convolutional neural network incorporating two innovative elements. The first element involves a trainable chromatic convolution layer that decomposes feature representations into six clinically relevant color channels namely white, red, light brown, dark brown, blue-gray, and black utilizing a differentiable soft-assignment matrix. This design enables the network to capture color patterns analogous to those used by dermatologists. The second component introduces a chromatic contrastive loss function that enhances lesion differentiation by explicitly clustering embeddings of lesions sharing similar color phenotypes while distancing those with dissimilar chromatic characteristics. In contrast to previous methodologies that consider color either invariant or as a fixed transformation, CANet adaptively identifies diagnostic color combinations pertinent to specific lesion categories. In the context of medical image analysis, specifically dermatological lesion classification, CANet attains a balanced accuracy of 92.7% on the HAM10000 dataset, compared to 87.6% for ResNet50 and 89.3% for EfficientNet-B4 under identical training conditions. Ablation experiments reveal that incorporating chromatic contrastive loss results in a 3.8% increase in accuracy, and the model’s learned color channels correspond closely with expert-annotated clinical color characteristics. CANet introduces an end-to-end framework that learns clinically grounded color representations, in contrast to approaches that treat color as either a fixed transformation or a nuisance variable.