UROP Proceedings 2020-21

School of Engineering Department of Computer Science and Engineering 114 Deep Learning for Ophthalmology Image Analysis Supervisor: CHEN Hao / CSE Student: CHEN Siyu / ISDN Course: UROP1000, Summer 2D fundus images and 3D optical coherence tomography (OCT) scanning volumes are two modalities of clinical data used for glaucoma grading. In this project, we use deep learning architecture to classify the samples of eye images into three categories according to visual features of these two modalities. The dual branch network was trained on data from 2D fundus images and 3D OCT images separately. The model can classify modality images into no glaucoma (without glaucoma are not samples without diseases, but images from patients with other eye diseases), early glaucoma, and moderate or advanced glaucoma. This classification methods can be an inspiration of eye disease diagnosis based on multi-modalities images using deep learning network. *The dataset and related materials used in this project are provided by ‘MICCAI2021 Contest: GAMMA’ which holds the copyright. Machine Learning in Healthcare Supervisor: CHEN Hao / CSE Student: KIM Hanbin / BIBU Course: UROP1000, Summer This report explores the application of machine learning in the healthcare industry, especially in drug discovery. Machine learning algorithms are being used in numerous aspects of drug discovery such as drug candidate selection and prioritization, and drug toxicity prediction. Through Computer-Aided Drug Discovery, pharmaceuticals could reduce R&D and clinical costs; Besides, an increasing trend of partnerships between Big Pharma and AI vendors could be seen. Among many aspects of drug discovery, this report investigates the use of machine learning in drug toxicity prediction by using Tox21 Challenge dataset and executing random forest model of a challenge winning team. This report also further gave a chance to think about the promising prospect of machine learning and its influence in drug discovery processes. Machine Learning in Healthcare Supervisor: CHEN Hao / CSE Student: SINGH Har Shwinder / PHYS Course: UROP1000, Summer Brain tumor detection is essential to identify brain cancer. This method aims to use semantic segmentation for AI tumor segmentation on MRI scans. A 3D Attention U-Net was used to train the 3D segmentation task of the BraTS 2021 Challenge Task 1. The Attention U-Net model improves the basic U-Net as the attention module increases focus on relevant feature maps, making the training more efficient. The model was trained with deep supervision, and data augmentation on the training dataset. Notably, positively labelled segmentation samples were cropped out as additional training samples. The highest results performance as submitted to the validation phase was: Dice of 0.817, 0.86 and 0.908 and Hausdorff 95% of 14.5, 9.9 and 5.97 for the enhancing tumor, tumor core and whole tumor respectively.