Explainable Multi-task Learning Approach for Skin Lesion Classification

The early diagnosis of skin cancer has significantly improved with the use of computer-aided techniques and deep learning (DL) models. However, existing methods often struggle with issues of interpretability and adaptability, which are crucial for clinical application. To address these limitations, we employed a Multi-Task Learning (MTL) approach that simultaneously performs classification and segmentation of skin lesions. This approach not only improves the accuracy and robustness of the models but also enhances their interpretability by incorporating Explainable AI (XAI) techniques into the MTL framework. Our convolutional-deconvolutional based MTL model, tested on the HAM-10000 dataset, demonstrated enhanced classification accuracy and interpretability with an Accuracy of 91.56% and IoU Score of 87.98%. The model outperformed state-of-the-art models, showing a marked enhancement in classification accuracy. Importantly, the use of MTL facilitates a reduction in model complexity while achieving superior performance, making this approach both powerful and efficient. From a practical standpoint, the replicability of our MTL framework is a key advantage, providing a scalable model for researchers and clinicians. The methodology's adaptability to different imaging datasets underscores its potential utility across various dermatological conditions. Future research could leverage this framework to further refine diagnostic accuracy in other complex imaging tasks, enhancing the scope of AI in medical diagnostics.