Improved CNN with BiLSTM model for early melanoma and skin lesion classification

Melanoma contributes to approximately 1% of all skin cancer cases but accounts for the majority of the deaths due to its aggressive behavior and high metastatic potential. Traditional diagnostic methods, including dermoscopy via the ABCDE rule, tend to be subjective and often lead to significant interobserver error, with accuracy rates generally ranging from 65% to 80%. This variability creates an urgent need for improved diagnostic tools that can enhance patient outcomes and reduce the burden on healthcare systems. Recent advancements in deep learning have transformed skin lesion classification. While CNNs excel at extracting spatial features, they often struggle to capture long-range contextual dependencies, an important factor in distinguishing visually similar lesions. Standard sequential models, such as Long Short-Term Memory (LSTM) networks, help address some of these gaps but have the drawback of being unidirectional. To overcome these limitations, this paper proposes a hybrid CNN-Bidirectional LSTM (BiLSTM) model. An all-encompassing preprocessing pipeline was developed using the HAM10000 dataset, which contains 10,015 dermoscopic images categorized into seven diagnostic labels. This pipeline includes noise reduction, lesion segmentation, data augmentation, normalization, and feature extraction, each aligned with the ABCDE rule. The CNN extracts hierarchical spatial information, while the subsequent BiLSTM layer learns contextual dependencies in both directions, allowing for a more nuanced description of lesions. The CNN-BiLSTM model with the suggested architecture has shown to be effective in the domain of automated classification of skin lesions, with an accuracy of 94.82%, a precision of 93.76%, a recall of 92.94% and an F1-score of 93.35% on the HAM10000 dataset, demonstrating its effectiveness in automated skin lesion classification with an inference time of only 80 ms per image, it is suitable for real-time clinical use. These results reinforce the clinical prospects of bidirectional contextual modeling in early melanoma detection and may have applications in broader computer-aided diagnosis systems.