Improving skin lesion classification through saliency-guided loss functions

Deep learning has significantly advanced computer-aided diagnosis, particularly in skin lesion classification. However, achieving high classification performance and providing explainable model predictions remain challenging in medical imaging. To tackle both performance and explainability challenges, we propose an effective method to enhance the performance of deep learning classifiers by integrating saliency scores directly into the loss function. Presuming that exploring various loss functions can significantly impact on the model performance, the proposed method is based on integrating a penalization weight derived from the saliency scores into the loss function, resulting a custom loss function for each XAI method. To evaluate the effectiveness of the proposed method, we have performed experiments on the challenging HAM10000 and PH2 datasets using the Inception-ResNet-v2, the EfficientNet-B3 and the ResNeXt classifiers for different XAI methods. The results demonstrate substantial enhancements compared to the baseline and relevant methods from the state-of-the-art. In fact, the proposed method achieved an accuracy of 94.3% and 98% for the HAM10000 and PH2 datasets, respectively, demonstrating an improvement over the standard loss function by 7% and 6% accuracy with an LRP-guided loss function. Thus, the designed integration improves model performance and reliability while implicitly assessing the effectiveness of the XAI techniques quantitatively through its ability to enhance classification.