Deep Learning-Based Melanoma Classification Using Hybrid DCNN-LSTM and DCNN-BiLSTM Architectures

Recent advancements in dermatology have improved the diagnosis and treatment of skin cancer. Early detection is particularly important, especially for melanoma, which is highly aggressive and can metastasize rapidly. However, diagnosing melanoma early is challenging due to its resemblance to atypical moles. While dermatologists assess lesions visually, microscopic examination is required for uncertain cases. The ABCD criteria are commonly used but may miss some melanomas. Dermoscopy offers greater accuracy and is preferred for reliable detection. AI and deep learning are revolutionizing dermatology, particularly in malignant melanoma diagnosis. This paper presents an AI system designed for accurate melanoma detection by incorporating a Deep Conventional Neural Network with transfer learning, data augmentation, and hybrid DCNN-LSTM and CNN-BiLSTM models. These techniques improve the performance of melanoma classification models. Experimental results demonstrate that the proposed methods surpass CNN approaches in accuracy, reaching 97.43%, specificity 99.6%, and, most importantly, sensitivity. This last metric, which represents the number of correctly identified malignant images, reaches 97.4% with the MobileNetV2-BiLSTM model.