Towards improved skin lesions classification using automatic hyperparameters selection and transfer learning

Classifying diverse skin lesions from dermoscopic images presents a challenge due to the striking visual similarities in size, shape, texture, and color across various lesion types. Complicating factors include artifacts in dermoscopic images, limited data availability, and the need for training deep architectures with millions of parameters, often resulting in overfitting and poor generalization. This study aims to construct a Computer-Aided Diagnosis (CAD) system capable of accurately classifying different skin lesions, leveraging both a publicly available ISIC 2019 challenge dataset and a private dataset.The research introduces three primary tasks. Task 1 focuses on enhancing data quality by addressing missing values, dermoscopic multi-resolutions, and imbalanced class distribution. Task 2 involves CAD development, employing a pretrained EfficientNet-L2 CNN architecture with self training technique (Noisy Student) as a feature extractor. supplementary metadata is incorporated using a dense neural network concatenated with the Convolutional Neural Network (CNN) output, followed by a classifier with eight units representing skin lesion classes and an unknown class. Task 3 centers on the creation of an automatic hyper-parameter selection system (CNN-AHPS) using MapReduce tool to identify optimal training hyperparameters for the combined architecture.Experimental results demonstrate that integrating metadata in the classification process enhances balanced multi-class accuracy and F1-score from 46.9% and 67.0% to 55.3% and 69.0%, respectively. Furthermore, CNN-AHPS surpasses default configurations, achieving a balanced multiclass accuracy and F1-score of 69.4% and 87.7%. the efficacy of NS-EfficientNet-L2 in classifying skin lesions is substantiated by experimental outcomes, outperforming state-of-the-art methods, including the winner of the ISIC 2019 challenge for skin lesion classification.