UROP Proceedings 2022-23

School of Engineering Department of Computer Science and Engineering 100 Deep Learning for Medical Image Analysis Supervisor: CHEN, Hao / CSE Student: FANG, Xiao / COSC Course: UROP2100, Fall This semester, I am extensively exposed to state-of-the-art deep learning technique used in medical image segmentation such as CNN, transformers, MLP permutator and even traditional feature extractor like Sobel operator. In detail, I propose 2.5D-Attentionunet, achieve breakthrough in combining MLP permutator and 2d CNN to solve anisotropic data problems in medical image, and also conduct experiments on four datasets (one private and three public) including lesion and organ segmentation using different network architecture. Results demonstrate that the work I have done achieves state-of-the-art performance on these datasets. Deep Learning for Medical Image Analysis Supervisor: CHEN, Hao / CSE Student: GU, Yi / COMP Course: UROP1000, Summer Detecting anomalies in medical images poses a challenge due to its high cost of annotating medical data. As a result, Anomaly detection requires neural networks to use normal data only and detect anomalies in the reference stage. Among the current methods in anomaly detection, the Student-Teacher (S-T) framework has demonstrated its effectiveness. However, prior works primarily focused on enhancing the output similarity among different models in normal data, with less emphasis on enabling models to behave differently in abnormal data. To address this limitation, we propose an explicit constraint for the spatial similarity of models’ features in the Student-Teacher framework. This approach can compel different models to learn similar normal features via distinct paths in the training phase, thereby promoting differences in abnormal data’s outputs. This method has been evaluated on the RSNA, VinDr-CXR, and Brain MRI medical datasets, demonstrating effective improvements on Autoencoder (AE) structure. Future experiments on S-T frameworks in industrial datasets are expected to test this method as a plug-and-play solution for most of Student-Teacher methods in anomaly detection.