Abstract:
Skin tumor segmentation is the preliminary step for medical diagnosis and consequent treatment of skin cancer. This writeup discusses our thesis work that utilizes deep CNN for skin tumor segmentation. The network uses down sampling and up sampling layers. The dense function or the fully connected layer has been used twice after the final down sampling in order to save computational time and system memory. To facilitate concatenation of matrices from the up sampling region and down sampling region, zero padding is provided during convolution and transpose convolution. Our network has depth comparatively lower than other contemporary deep neural networks used for image segmentation. In spite of that our model has performed well with satisfactory accuracy of 55% after 5 epochs and 62.23% after 10 epochs. It is worth mentioning that our model was trained with 1000 images, tested on 1000 and validated on 600 medical images containing(or not) skin tumors
Description:
Supervised by
Prof. Dr. Md. Ruhul Amin
Professor,
Department of Electrical and Electronic Engineering,
Islamic University of Technology.
