Early Diagnosis of Melanoma by Augmenting Feature Extraction of Epidermis using Faster Region-Based Convolutional Neural Networks

As fatal diseases like cancer are becoming more common nowadays and skin cancer being one of them, detection of the disease and its treatment at an early stage holds vital importance in order for the patient to recover from it both physically and mentally. Similarly, number of melanoma cases have been increasing gradually and subsequently. Introducing advanced and state-of-the-art technologies like computer vision to the medical field has proven nothing but advantageous to the patients as well as doctors. Our study helps identify skin cancers like melanoma without getting a physical biopsy done. The affected region of the skin is given as an input to our model and which runs inference on it to identify whether melanoma is detected or not. Image segmentation and feature extraction of the image takes place and then the Fast RCNN R101-FPN model from detectron model zoo is used for image classification. Detectron2 consists a number of object detection algorithms like faster R-CNN, Densepose, RetinaNet that help us detect the problem region. The model consists a number of hyper-parameters and the input images of lesion is up-sampled to give the desired output. Faster R-CNN has been used for instance detection in this research.