An Optimized Deep Learning framework for Skin Cancer Classification with Hybrid CNN Architecture and Data Augmentation

Abstract

Introduction:  Skin cancer, especially melanoma, remains an evolving international public health concern as it advances in an aggressive manner and the rates of its occurrence are rising. It is crucial to identify and treat skin lesions on time and with accuracy to improve patient outcomes. This work provides an improved deep learning pipeline for automatic skin lesion classification using CNN, integrated with optimized data pre-treatment and augmentation methods. The HAM10000 dataset, consisting of 10,015 dermatoscopic images in seven diagnostic classes, is considered the main dataset. The model pipeline includes advanced steps like using dull-half razor filtering to reduce hair interference, segmenting lesions with autoencoders, and balancing the classes through under-sampling and over-sampling. Using transfer learning methods, different pre-trained CNN models like DenseNet169, ResNet50, InceptionV3 and VGG16, are compared based on their accuracy, precision, recall, and F1-score. Numerical results show that the DenseNet169 has a better performance when applying the under-sampling process, while the ResNet50 yields better performance when it uses the over-sampling process. An ensemble model that utilizes the best aspects of these architectures is introduced, which obtains an expected accuracy of over 95%, better than the benchmark VGG16 and DenseNet161.The discovered network also confirms the importance of patient-specific deep learning models, and (patient-specific) pre-processing pipelines are applied in dermatological diagnostics, leading to generally feasible AI-assisted CAD systems to support clinicians during real-world decision-making.