Melanoma Detection Using Machine Learning

Abstract

Pulmonary Melanoma is the leading cause of cancer-related deaths. Early detection of lung cancer can significantly reduce the mortality rate. The main goal of this thesis is to segment the boundaries of lesions from Computed Tomographic images to assist the pulmonologists so that decisions can be made easier for them. The dataset is taken from Medical Decathlon. The dataset contains pre-labeled tumor images annotated by medical experts to enrich the level of analysis. The automated models used for the segmentation of images are: Nested U-NET, Mask RCNN, and DeepLabV3, and an Ensemble of these models is created. Pytorch and FastAl libraries were used for training the dataset. 33% of images were used for validation. The learning rate was at least during the data to avoid overfitting. The metric used for calculating results was Dice. The individual Dice scores of Nested U-Net, Mask RCNN, and DeepLabV3 are 0.7487, 0.7811, and .882 respectively and the accuracy for the ensemble is 0.895, and state-of-the-art is 0.77.