Automatic Segmentation and Classification of Skin Cancer Cells using Thresholding and Deep Learning based Techniques

The proposed manuscript derives its origin from the fact that annually about 57 million people die due to skin cancer. Skin biopsy is the most widely practiced method to detect skin cancer. Fear of bleeding, and infection raises the concern of patients to agree for skin biopsy. Also, its' high cost, and requirement of specific infrastructure is a limitation in its use. The proven potential of artificial intelligence motivated us to resolve the above-mentioned challenges and propose an accurate diagnosis method for skin cancer. In this research, we apply different preprocessing methods for removing rulers and hair from the images in the dataset. Also, we employ different segmentation techniques, such as color based segmentation, thresholding, edge detection, and region-based segmentation to segment skin cancer images. Next, we compare the performance of these techniques by implementing them on the same testing dataset. We use PSNR, SSIM, entropy, and brisque score to evaluate the quality of segmented images. The mean entropy of 0.52 bits/pixels, mean PSNR of 38.75dB, mean SSIM of 0.82 and brisque score of 119.09 on the testing dataset prove the efficacy of the proposed model. Further, finetuning the VGG-16 model for classifying the testing dataset comprising 617 images improve the performance and reported 98% accuracy. The high accuracy and good quality of segmented images justify the performance of the model.