School of Engineering Department of Electronic and Computer Engineering 164 Visual-Language Large Foundation Models and Their Applications in Medical Image Analysis Supervisor: LI Xiaomeng / ECE Student: WANG Yueming / COMP Course: UROP 1000, Summer This report details the progress of a research project focused on applying deep learning techniques to two distinct medical imaging challenges: skin lesion classification and spleen CT segmentation. For the first task, a ResNet-50 based classification network was successfully implemented to differentiate between benign and malignant skin lesions from the ISIC 2016 dataset. By employing techniques such as data augmentation and weighted random sampling to address class imbalance, the model achieved a final test accuracy of 86.28% and an AUC of 0.8455, satisfying the project requirements. The second task involved developing a 3D U-Net for segmenting the spleen from abdominal CT scans. While the model training demonstrated convergence, the initial predictions were inverted, with the model segmenting the background instead of the organ. This report analyzes this issue, presents the current progress, and outlines the corrective steps needed to validate the model’s performance on spleen segmentation. Visual-Language Large Foundation Models and Their Applications in Medical Image Analysis Supervisor: LI Xiaomeng / ECE Student: ZHANG Li / COMP Course: UROP 1100, Fall Medical image classification and segmentation is an important branch of the field of computer vision. In the project, I have learned about the basic logic of machine learning, along with the code implementation using PyTorch, including the transforms, data-loader, loss function and optimizer, model architecture, trainingvalidation process, and evaluation, as well as training result demonstrating using Tensorboard. Then, I used what I had learned to tackle two main problems, skin lesion classification, which is a 2D image classification task based on pre-trained resnet50, and spleen CT segmentation, a 3D image segmentation task using UNet structure. I have also learned to set up environments with Anaconda, train on a remote GPU server, and use GitHub Copilot wisely to help me learn the syntaxes and implement the codes more efficiently and effectively. DESIGN OF INTEGRATED AMPLIFIERS Supervisor: LUONG Howard Cam / ECE Student: REN Junru / ELEC Course: UROP 1100, Summer This report gives a brief summary of the feedback principles in Behzad Razavi’s Design of Analog CMOS Integrated Circuits. It focuses on four standard feedback topologies: voltage-voltage, current-voltage, voltage-current, and current-current. Theoretical synthesis and circuit-level insights are used to look at how each topology affects input/output impedance, gain, and real-world use. Loop gain calculation, loading effects of feedback networks, and noise considerations are some of the most important ideas. The main focus is on how feedback mechanisms systematically improve circuit stability and performance. The analysis includes mathematical formulas from the textbook to show how important feedback is to analog design.
RkJQdWJsaXNoZXIy NDk5Njg=