Hybrid Deep Learning Models for Skin Lesion Classification: A Comparative Review and Future Directions

Abstract

The accurate and early characterization of skin lesions is crucial to the timely intervention in the diagnosis of skin cancers, especially melanoma. Due to the penetration of artificial intelligence (AI) in medical imaging, deep learning techniques particularly hybrid models that integrate Convolutional Neural Networks (CNNs) with attention mechanisms, transformers and sequential networks such as long short term memory (LSTM) have demonstrated promising progress for improving classification performance. In this paper, we provide a thorough survey on recent hybrid deep learning architectures proposed for skin lesion classification purpose with a focus on methods, employed datasets and comparative results. We perform a systematic review of key works to expose predominant challenges, such as class imbalance, low interpretability and restricted generalisability across lesion types. We also point out what need to be improved and introduce a new concept hybrid model combining EfficientNet-B6 and LSTM for this requirement. Comparative analysis if the existing benchmark comparisons are provided to justify the divergence. Recommendations for robust and interpretable AI system in dermatologia have been discussed at the end of the paper.

Ähnliche Arbeiten