Analyzing the Diagnostic Efficacy of Deep Vision Networks for Malignant Skin Lesion Recognition

Abstract

Cancerous skin lesions affect millions of people worldwide each year and is one of the most treatable forms of cancer. While intensive biopsy testing and processing is required to confirm the presence of a malignant skin lesion, the detection of skin lesions by dermatologists on the primary level has always been based upon visual markers and sight-based perception based upon a defined set of diagnostic rules. The automation of this classification process has been achieved in the past for traditional machine learning algorithms and novel deep networks but faces challenges when the diagnosis is performed upon images of varied illumination and spatial orientation. This paper proposes a novel ensemble approach towards skin lesion classification by employing transfer learned pretrained deep learning image networks for the automated diagnosis process. Popular ImageNet Trained Networks such as DenseNet, Inception ResNetV2, VGG16 and MobileNet have been individually fine-tuned, tested and evaluated for identifying the type of skin lesion. A final integration of the best ensemble combination was performed based upon a search- based strategy to find the optimal combination for maximal reliability. The system was tested against benchmark datasets including HAM1000 and ISIC, showcasing an accuracy of 90%, precision of 0.895, and recall of 0.89 and the proposed combinational network showcases significantly better results than several existent state of the art skin cancer classification models in terms of accuracy, precision and recall.