Project information
- Associated With: EEE 402: Artificial Intelligence & Machine Learning Laboratory
- Project date: February, 2024
- Project Report: Click to View in IEEE format
- Skiils: : Convolutional Neural Networks (CNN) · Deep Learning · TransUnet · Medical Image Segmentation
Project Details:
Our study aims to enhance neural network segmentation for medical imaging tasks focused on single roughly elliptically distributed objects. Our method involves training neural networks on polar transformations of the original dataset, ensuring alignment of the polar origin with the object’s center point. This innovative approach effectively reduces dimensionality and facilitates the separation of segmentation and localization tasks, thereby promoting network convergence. We explore two distinct strategies for determining the optimal polar origin: leveraging segmentation trained on non-polar images and utilizing a model trained to predict the optimal origin. Using the ISIC-2018 dataset, we replicate and extend upon a method proposed in a prior study. Additionally, we conduct a comprehensive comparison with the TransUNet model and introduce a novel architecture combining TransUNet and stacked hourglass models.