Classification of Lesion Images Using Transfer Learning Approach

Abstract

Caused due to the uninhibited increase of abnormal skin cells, skin cancer is a result of unrepaired DNA damage to skin cells which in turn, leads to mutations, or genetic defects. These defects cause rapid multiplication of skin cells and they eventually formulate malignant tumours. Although skin cancer is one of the most lethal types of cancer, a fast diagnosis can lead to a very high chance of survival. The diagnosis of skin cancer is primarily performed using visual methods, usually an initial clinical screening. Dermoscopic analysis, a biopsy and histopathological examination consist of the conventional methods that follow the clinical screening. An automated classification of skin disease using images is a difficult job due to the microscopic variability of the appearance of different classes of skin lesions. Over the last few years, convolutional neural networks (CNN) have been increasingly employed for the task of automatic and semi-automatic image classification. Through this work, we aim to use a transfer learning-based deep learning approach to detect cancerous lesions in dermatological images. The process would involve pre-processing and data augmentation tasks being performed on the lesion images. Following this, a pre-trained transfer learning model would be fine-tuned and used for feature-selection and a classifier model would be added on top of it to classify the images of skin lesions into ‘malignant’ and ‘benign’ categories. The model was tested using standard evaluation metrics to evaluate its effectiveness. Our results show that a transfer learning approach can work as an effective screening tool to detect cancerous lesions.