Deep Pixel-wise supervision and deep mask pixel-wise supervision for skin lesion classification

Background: Utilizing automated systems for diagnosing malignant skin lesions promises to improve the early detection of skin diseases and increase patients’ survival rates. However, current classification methods primarily focus on global features, often overlooking local patterns.

Methods: We apply two training approaches to address this limitation: deep pixel-wise supervision using a constant map and deep mask pixel-wise supervision employing a segmentation mask. Both methods apply supervision to each pixel of the feature map of the network, providing guidance to hidden layers and encouraging the learning of reliable patterns for class distinction. Combining deep and pixel-wise supervision effectively directs the network’s attention to lesions with high precision.

Results: Our experiments, spanning binary and multiclass classification using ISIC and PH2 datasets, reveal enhanced accuracy and sensitivity. The proposed model achieves 90.7% and 90.5% accuracy for binary classification on ISIC 2017 and PH2 datasets, respectively, and 88% accuracy for nine-class classification on a combined ISIC 2019 and 2020 dataset.

Conclusion: Comparative evaluations demonstrate superior performance compared to state-of-the-art methods, validating the effectiveness of deep pixel-wise supervision for skin lesion classification.

Full text available here

Preprint version available here

Trained models can be downloaded from here