School of Engineering Department of Computer Science and Engineering 95 Deep Learning for Medical Image Analysis Supervisor: CHEN Hao / CSE Student: PARK Juwon / COMP Course: UROP 1100, Spring This report focuses on the UROP 1100 journey regarding research on deep learning foundations and applying its mechanisms to image segmentation. The field of computer science is utilized in various fields, and one of the most influential developments in recent years is the Convolutional Neural Network. This system model allowed researchers to train the model for various tasks related to finding patterns, identifying, and recognizing objects. A widely used example of CNN is in the field of bio-imaging, a method to process and visualize images of biological objects or systems. While various imaging techniques were invented as early as the late 19th century, the method to visualize and detect different objects from the image became more accurate and reliable in recent years. This paper will focus on the fundamentals of the Convolutional Neural Network, and its implications in the field of 2D segmentation. Deep Learning for Medical Image Analysis Supervisor: CHEN Hao / CSE Student: WU Tong / DSCT Course: UROP 1100, Summer The concept bottleneck model first predicts concepts provided in the course of training and then uses these concepts to predict the label. This model allows for interventions by altering the predicted concept values, which in turn affect the final prediction outcome. This study aims to validate these conclusions through computational implementation. By running the code provided in the original paper written by Koh et al. in 2020, we replicate key experiments and analyze the results. Our findings from this research confirm that concept bottleneck models achieve comparable task accuracy to standard end-to-end models while having high concept accuracy. Diffusion Models for Medical Imaging and Analysis Supervisor: CHEN Hao / CSE Student: LOH Angus Han Jern / COMP Course: UROP 1100, Spring The survival rate of patients with breast cancer can be significantly reduced via early detection. Early detection can be conducted with mammograms, and thus can drastically reduce treatment costs. Tumours can be detected through segmentation techniques, as it helps in image analysis, detection, feature extraction, classification and treatment. The segmentation technique used in the experiment is semantic segmentation, notably U-Net segmentation. Compared to other deep learning models, U-Net does not require many annotated images and is suitable in the presence of high-performance GPU computing. In this report, a detailed breakdown of the U-Net model used will be provided, including the training and outcome metrics to evaluate the performance of the model.
RkJQdWJsaXNoZXIy NDk5Njg=