School of Engineering Department of Computer Science and Engineering 104 Deep Learning for Medical Image Analysis Supervisor: CHEN Hao / CSE Student: LI Chengxi / COMP Course: UROP 1100, Fall CNNs, Vision Transformers (ViTs), and Swin Transformers have made great progress in computer vision but often struggle with overfitting due to spatial inductive biases in their architectures. To solve this, we propose the Pointer model, which avoids patching or window-based methods and applies self-attention directly to the entire image. This helps capture both local and global features while preserving spatial continuity. The Pointer model uses a U-Net-like structure with Transformer blocks for feature extraction. A new Pointer Attention mechanism reduces spatial biases by computing attention across the whole image. However, the model faces challenges in learning positional information, and we are exploring methods like row-column encoding to address this. We are continuing to improve positional encoding and overall performance, aiming to make the Pointer model effective for image segmentation tasks. Deep Learning for Medical Image Analysis Supervisor: CHEN Hao / CSE Student: LIU Runsheng / COMP Course: UROP 3100, Fall Computational cytology is an essential, rapid-developing topic in the field of medical image analysis and disease diagnosis. Analyzing digitized cytological Whole Slide Images (WSI) by computer-aided technologies for WSI-level classification is yet challenging due to huge size -usually gigapixel- of WSIs, imbalance and ambiguity among specimen of different types, laborious process of labeling, etc. One strategy is conducting Region of Interest (RoI) selection and classification based on cell detection and comprehensive learning. However, this kind of cell-level methods acquires tedious manual screening and labeling the cells, which is of huge amount even in a single WSI. On the other hand, Multiple Instance Learning (MIL) is a currently very effective and well-developed deep learning strategy in the field of histology WSI classification, due to its applicability on gigapixel images and effectiveness on feature extraction and interpretability. Naturally, MIL is adopted to the classification task on cytological WSIs and get rid of the need of cell-level labeling. In our project, we verify the effectiveness, and explore the reasonability of MIL on cytological WSI classification tasks regarding to its feature learning and interpretability. Experiments show that attention-based MIL lack reasonable both utilization of the extracted features, and interpretability. We discuss possible reasons of it, and raise solutions towards this task persisting the use of MIL. We will also report what is the current progress of our exploration and development of our new ideas.
RkJQdWJsaXNoZXIy NDk5Njg=