Temporal integration of ResNet features with LSTM for enhanced skin lesion classification

Abstract

• Proposed the R-LSTM50 architecture, combining ResNet50 for feature extraction with LSTM for temporal modeling, to improve skin lesion classification accuracy. • Achieved superior performance on ISIC2020 and HAM10000 datasets, with 95.72% and 94.23% accuracy, respectively, outperforming existing transfer learning models. • Evaluated the proposed model using multiple metrics, including F1-measure, sensitivity, and specificity, ensuring comprehensive performance validation. • Designed the R-LSTM50 model to be deployable across diverse hardware and software platforms, enabling its use as a pioneering medical diagnostic tool. The precise classification of skin lesions is essential for the early identification and efficient treatment of skin cancer. This research combines ResNet-50 for spatial feature extraction with Long Short-Term Memory (LSTM) networks for temporal analysis and introduces an innovative hybrid deep learning model, Residual network-Long Short-Term Memory (R-LSTM50). The approach tackles issues including class imbalance and the necessity for enhanced sensitivity in identifying malignant tumors. The ISIC2020 and HAM10000 benchmark datasets were employed for evaluation, utilizing sophisticated data augmentation methods and weighted loss functions to improve performance. Evaluation measures such as accuracy, sensitivity, specificity, and F1-score were employed to verify the model. Experimental findings indicate that R-LSTM50 attains state-of-the-art performance, achieving accuracies of 95.72% and 94.23%, F1-scores of 92.78% and 93.46%, sensitivities of 95.24% and 92.41%, and specificities of 92.08% and 90.33% on ISIC2020 and HAM10000, respectively. The results demonstrate the robustness and clinical significance of R-LSTM50, confirming its reliability as a tool for automated skin lesion classification. The hybrid design and sophisticated preprocessing techniques enhance the management of class imbalance and intricate feature interactions, establishing the proposed model as a notable improvement over current methodologies.