UROP Proceedings 2022-23

School of Engineering Department of Computer Science and Engineering 102 Deep Learning for Medical Image Analysis Supervisor: CHEN, Hao / CSE Student: XU, Jiayi / SENG Course: UROP1100, Fall Diabetic retinopathy (DR) and diabetic macular edema (DME) are leading causes of blindness in the workingage population in most countries. Since automatic grading of DR and DME helps ophthalmologists diagnose diseases more accurately. It plays an important role in clinical practice and more and more people began to pay attention to and study this subject. In the first two months of this semester, I learned some fundamental structure of deep learning neural network, studied the paper proposed by Li. which makes a huge contribution to jointly grading DR and DME, and finally reproduced the model by myself. After finishing this basic training, I started to grope for multi-modality direction. In this report, I try to unscramble the paper relevant to DR grading in my words based on my own understanding and then illustrate some progress in multi-modality exploration. Deep Learning for Medical Image Analysis Supervisor: CHEN, Hao / CSE Student: ZHANG, Rushan / AE Course: UROP3100, Spring UROP4100, Summer The diagnosis based on whole slide images of lymph nodes is of great significance for the identification of breast cancers. However, diagnosis by pathologists is time consuming and error prone due to the large size of the whole slide and the small size of the lesion regions. In this project, we propose to introduce weeklysupervised detection modules to guide the multiple instance learning algorithm for whole slide image classification. To be more specific, we train a detector to identify regions of interests in low resolution images, then crop out the regions of interests from the high resolution images and feed them into the MIL algorithm. The proposed algorithm is implemented and experimented on the CAMELYON16 dataset. Deep Learning for Medical Image Analysis Supervisor: CHEN, Hao / CSE Student: ZHU, Zhengjie / COMP Course: UROP1100, Fall Whole Slide Image (WSI) classification of histopathological tissues differs significantly from natural image classification. WSI has a gigapixel resolution and usually lacks pixel-level annotation. When only slide-level labels exist, weakly supervised multiple instance learning (MIL) is usually employed to handle WSI classification. However, existing MIL methods still face two major problems: 1. Cannot extract critical information well on patches 2. It is hard to effectively derive the classification of bag labels from a large number of instances (patches) in a bag (WSI). To solve these two problems, first, I designed a self-supervised pretext task based on contrastive learning, which utilizes multi-resolution contextual clues of Whole Slide Image to extract features rich in semantics. Secondly, I will introduce an explicit data distribution modeling method based on clustering, and an attention-based MIL aggregator to classify WSI-level labels accurately, named CAMIL. I will conduct experiments on two classic public data sets with WSI labels, the CAMELYON16 dataset and the TCGA lung cancer dataset, to demonstrate the effectiveness of this method.