UROP Proceedings 2021-22

School of Engineering Department of Computer Science and Engineering 100 Machine Learning in Healthcare Supervisor: CHEN Hao / CSE Student: KHAN Asif / COMP Course: UROP1100, Spring In light of the recent developments in artificial intelligence and medical technologies, researchers all around the world have been exploring and implementing various machine learning methods geared toward healthcare applications. A field that is most commonly researched is medical image processing. Machine learning can be utilized by researchers and doctors to optimize and provide more accurate diagnoses through medical image analysis, and several architectures and algorithms have been designed to facilitate effective image processing techniques. An initial starting point for further boosting medical research would be to start through the method of nuclei detection. With the help of the datasets provided by Kaggle’s 2018 Data Science Bowl, this report would build and evaluate several networks for image segmentation. This report would also explore the potential prospects of machine learning and how it could effectively be used in healthcare applications. Machine Learning in Healthcare Supervisor: CHEN Hao / CSE Student: SINGH Har Shwinder / PHYS Course: UROP1100, Fall This paper proposes a Deeply Supervised Attention U-Net Deep Learning network with a novel image mining augmentation method to segment brain tumors in MR images. The network was trained on the 3D segmentation task of the BraTS2021 Challenge Task 1. The Attention U-Net model improves upon the original U-Net by increasing focus on relevant feature maps, increasing training efficiency and increasing model performance. Notably, a novel data augmentation technique termed Positive Mining was applied. This technique crops out randomly scaled, positively labelled training samples and adds them to the training pipeline. This can effectively increase the discriminative ability of the Network to identify a tumor and use tumor featurespecific attention maps. The metrics used to train and validate the network were the Dice coefficient and the Hausdorff metric. The best performance on the online final dataset with the aforementioned network and augmentation technique was: Dice Scores of 0.858, 0.869 and 0.913 and Hausdorff Distance of 12.7, 16.9 and 5.43 for the Enhancing Tumor (ET), Tumor Core (TC) and Whole Tumor (WT).