Project Description

This project focuses on the development and evaluation of a deep learning model for segmenting brain MRI images, with a primary goal of accurately identifying tumor regions in patients with lower-grade glioma (LGG). It combines foundational CNN concepts with advanced biomedical image segmentation techniques using the U-Net architecture.

The project begins with environment setup and basic CNN training on the CIFAR-10 dataset to establish a solid understanding of convolutional networks. It then progresses to transfer learning using ResNet18 before transitioning into the core task of biomedical image segmentation.

The U-Net model is implemented and trained on a publicly available LGG MRI dataset from Kaggle. Key evaluation metrics such as Intersection over Union (IOU) and Dice coefficient are used to assess model performance. Additional experiments include data augmentation, learning rate scheduling, and loss function comparison (e.g., Binary Cross-Entropy vs. Focal Loss).

Project Goals:

1. Develop a Robust Segmentation Pipeline
   Design and implement a U-Net-based deep learning pipeline capable of accurately segmenting brain tumors from MRI scans, with an emphasis on modularity, reproducibility, and scalability.
   
   

2. Achieve High Segmentation Accuracy
   Optimize the model using techniques like data augmentation, learning rate scheduling, and loss function tuning to achieve at least a Dice score > 0.85 and IOU > 0.80 on the test set.
   
   

3. Compare and Analyze Model Variants
   Evaluate the performance of advanced model variants (e.g., Attention U-Net, MedSAM) and compare them against the baseline U-Net to understand trade-offs in accuracy, training time, and generalization.

Team Members