School of Engineering Department of Computer Science and Engineering 103 Deep Learning for Medical Image Analysis Supervisor: CHEN Hao / CSE Student: KWAK Dowon / BIEN Course: UROP 1000, Summer The rise of deep learning in computational pathology (CPath) has enabled impressive progress in automating the analysis of whole slide images (WSIs) across various diagnostic tasks. However, real-world deployment remains a challenge due to distribution shift, which refers to systematic discrepancies between the source distribution ( , ) and the target distribution ( , ), where denotes input WSIs and their labels. Common sources of distribution shift in CPath include differences in tissue staining protocols, scanner hardware, imaging resolution, and institutional practices. Distribution shift appears in various forms such as covariate shift � ( ) ≠ ( )�, label shift ( ( ) ≠ ( )), or conditional shift ( ( | ) ≠ ( | )). In CPath, subtle visual heterogeneity in color, texture, and spatial tissue patterns often translate to significant shifts in embedding space leading to poor generalization of models on unseen target domains. To address these challenges, test-time adaptation (TTA) was proposed as a practical solution that adapts models to unseen target domains during inference. TTA methods update model parameters or decision strategies online without access to any target domain labele. Despite its versatility and efficiency, TTA methods for CPath are extremely understudied and face several limitations. Entropy-minimizing approaches can reinforce incorrect predictions when distributional uncertainty is high. Also, many methods operate persample and discard potentially valuable information observed earlier. Techniques relying on ensemble models, test-time augmentations, or iterative optimization introduce latency incompatible with clinical workflows. To overcome these drawbacks, I propose CASH: Cache-Aware Shift Handling, a training-free and memoryaugmented TTA framework tailored for CPath. CASH maintains a class-conditional memory cache of test embeddings and employs slide-level attenuation algorithm to discourage overconfident predictions. CASH accumulates a highly informative and domain-invariant representations of relevant classes on-the-fly and reinforce model decisions accordingly in computationally efficient manner. Deep Learning for Medical Image Analysis Supervisor: CHEN Hao / CSE Student: LEE Hyunjin / COMP Course: UROP 1100, Summer Convolutional Neural Networks (CNNs) play a crucial role in modern medical image analysis, particularly in image segmentation tasks. This paper investigates the application of the U-Net architecture to two distinct biomedical segmentation tasks: retinal vessel segmentation and dental structure segmentation. Both tasks utilized a similar U-Net configuration with minor adjustments, and both resulted in a dice coefficient of greater than 0.75 with minimal computation time, however the model was less effective for the retinal vessel images compared to the dental images. These results demonstrate not only the versatility and effectiveness of U-Net in handling diverse medical imaging challenges but also highlight the impact of dataset size and anatomical consistency on model performance.
RkJQdWJsaXNoZXIy NDk5Njg=