School of Engineering Department of Computer Science and Engineering 105 Deep Learning for Medical Image Analysis Supervisor: CHEN Hao / CSE Student: OH Yeontaek / COMP Course: UROP 1100, Summer This report will discuss deep learning models that can automatically diagnose breast cancer using mammography. Two types of architectures are reviewed, which are CNN-based and transformer-based architectures on the CBIS-DDSM dataset, which provides the full mammograms along with the corresponding ROI (Region of Interest) paths and cropped images. The CNN-based models displayed stronger initial performance, while the transformer-based models faced a few challenges, such as image distortion and loss of details from resizing. Faster R-CNN, DETR, ViT, and Two-Branch Transformer models are implemented and trained using the CBIS-DDSM dataset on ‘Kaggle’. A test dataset, which was separated from the training dataset, was fed to check the accuracy and loss for evaluation. Deep Learning for Medical Image Analysis Supervisor: CHEN Hao / CSE Student: PARK Juwon / COMP Course: UROP 2100, Spring UROP 3100, Summer Over the past two months, significant progress has been made on the development of a diffusion model aimed at developing dental implant prosthetics from 3D dental data. The previous process involved 3D segmentation and mesh generation from CBCT scans and learning about diffusion models and point-cloud generation. Upon learning about the diffusion model, a curated dataset was required for training the model. The raw dataset, initially consisting of numerous patients’ dental mesh files, required manual cleaning, verification, quality control and formatting. In addition, adaptation to an existing diffusion model into our scope and combining all the raw datasets into a single compatible format have to be done. The nature of medical-imaging and medical-AI models requires extensive amounts of data for the model to be robust and effective for various medical conditions. Consequently, a significant amount of time was spent during the UROP period for data cleaning. This report will focus on the progress being made during the data cleaning and diffusion model implementation process, and will discuss the further plan for different attributes and configuration changes which will be made for better model performance. Deep Learning for Medical Image Analysis Supervisor: CHEN Hao / CSE Student: TSE Ling Kun / COMP Course: UROP 1100, Spring This report details progress in developing cytology image analysis frameworks in early stages: 1) Systematic literature review and methodology categorization, 2) Comprehensive data preparation of APACS23 and Oral2021 datasets with customized COCO conversion strategies, and 3) Baseline experiments using MMDetection frameworks. The data processing phase resolved critical annotation discrepancies through center-point sampling for overlapping cells and adaptive mask conversion algorithms. Three instance segmentation models (Mask R-CNN, Cascade Mask R-CNN, MS R-CNN) were implemented on Oral2021 data, establishing performance benchmarks. More experiments of models on Oral2021 and other datasets will be conducted in the future. Current results demonstrate 25%~35% mAP scores, and a seemingly erroneous result is found in Mask Scoring R-CNN’s result, which should be further investigated.
RkJQdWJsaXNoZXIy NDk5Njg=