School of Engineering Department of Computer Science and Engineering 104 Data-Efficient, Domain Generalizable and Interpretable Deep Learning Supervisor: CHEN Hao / CSE Student: WANG Zeyu / COMP Course: UROP1100, Fall Deep learning models have shown great potential in detecting mitotic figures in histopathology images, the density of which is known to be highly related to tumor proliferation and thus crucial in diagnosing cancer. However, performance of a model would degrade significantly if images for testing are from different laboratories than the images for training. This is caused by domain gaps due to difference in specimen acquisition methods. Under hypothesis that the whole slide image(WSI) scanner device plays a decisive role, the MICCAI-MIDOG challenge is organized to seek for machine learning solutions that are invariant to the domain shift. In this UROP project, I developed augmentation-adversarial-based pipeline to resolve the domain gaps. Data-Efficient, Domain Generalizable and Interpretable Deep Learning Supervisor: CHEN Hao / CSE Student: ZHANG Yuhao / SENG Course: UROP1000, Summer Diabetic Retinopathy (DR), often occurs with diabetes mellitus. It is a leading cause of blindness for many patients. Automatic grading may help the patient receive the healing earlier. This report will study and discuss several different ways of using machine learning to grade Diabetic Retinopathy automatically. Among all those models, I discovered that the Convolutional Attention Block plays an important role in grading Diabetic Retinopathy. By using several different models which used the Convolutional Attention Block and changing the hyperparameters, this report tries to discover a way to grade Diabetic Retinopathy by using machine learning methods that reaches a high accuracy.