School of Engineering Department of Electronic and Computer Engineering 160 Artificial Intelligence Methods for Medical Videos Supervisor: LI Xiaomeng / ECE Student: APHIMOOKKUL Nattgarni / COMP Course: UROP 1100, Spring In this semester, I have read papers in various topics such as large language model (LLM) and Retrieval Augmented Generation (RAG) in the medical field. Some of the papers are from the PhD students under Dr. Li. Also, I have done and tested some simple personal projects to utilize on what I have learned. Federated Learning with Medical Images Supervisor: LI Xiaomeng / ECE Student: GUO Yu / COMP Course: UROP 1000, Summer In the modern age, many tasks cannot be handled simply by rule-based systems. There are numerous rules and guidelines that cannot be implemented by human power. Contemporary approaches are data modeling, in other words, using a number of parameters and functions to find the most resembling representation of the given dataset. The process of automatically fitting the numeric representation of the dataset is so called Machine Learning. This report summarizes the process of learning Machine Learning with PyTorch. Starting from the basic operations of data (Tensors) with PyTorch and the root principles of Machine Learning, the report covers the workflow of data post-processing, neural network building, testing, training and evaluating a model using the given medical image datasets. Federated Learning with Medical Images Supervisor: LI Xiaomeng / ECE Student: VIJ Saumik / MATH-CS Course: UROP 1000, Summer This report explores deep learning applications in medical image analysis through two projects: skin lesion classification and spleen CT segmentation. A ResNet-inspired convolutional neural network (CNN) was developed from scratch to classify dermoscopic images from the ISIC 2016 dataset as either benign or malignant, achieving a test accuracy of 63%. A 3D U-Net model was implemented for spleen segmentation from CT scans (Medical Segmentation Decathlon dataset), yielding a low test Dice score of 0.0094 due to limited data and model complexity. These projects highlight deep learning’s potential and challenges in automating diagnostics, offering insights into CNN design, data preprocessing, and performance optimisation for healthcare applications.
RkJQdWJsaXNoZXIy NDk5Njg=