diptarup794 / mri-image-segmentation Goto Github PK

MRI Tumor Segmentation with UNET and Res-UNET

Table of Contents: Project Overview Data Source Model Architecture Evaluation Metrics Results Installation and Setup Usage Contributing License

Project Overview: This project aims to perform image segmentation on a dataset of MRI scans from The Cancer Imaging Archive (TCIA). The goal is to detect and segment brain tumors within the MRI images. We employed two popular segmentation models, UNET and Res-UNET, to achieve this task.

The evaluation was carried out by calculating the Dice Loss and Jaccard Loss for both models, allowing for a comparative analysis of their performance. This approach enables accurate detection and segmentation of brain tumors, providing valuable insights into their location and relative size.

Data Source: The dataset used for this project consists of real MRI scans of patients sourced from The Cancer Imaging Archive (TCIA). The dataset contains labeled MRI images, with annotations indicating the presence and location of brain tumors.

Model Architecture: UNET: A popular convolutional neural network architecture designed for image segmentation. It features a symmetric "U" shape with a contracting path (encoder) and an expanding path (decoder). The skip connections in UNET enable the model to maintain high-resolution features. Res-UNET: A variant of UNET that incorporates residual connections. These connections allow for better gradient flow during training, potentially improving convergence and reducing overfitting. Evaluation Metrics Dice Loss: A metric used to evaluate the overlap between the predicted segmentation and the ground truth. It is derived from the Dice Coefficient, which measures the harmonic mean of precision and recall. Jaccard Loss: Based on the Jaccard Index (also known as Intersection over Union), it quantifies the similarity between the predicted and actual segmentations.

Results: UNET: The Dice Loss for UNET was calculated as 0.01130, indicating a high degree of accuracy in segmenting the brain tumors. Res-UNET: The Dice Loss for Res-UNET was 0.01140, showing slightly lower performance than UNET. Jaccard Loss: Although not specifically mentioned, this loss function complements Dice Loss, providing further insights into model accuracy. These results suggest that both models are effective for brain tumor segmentation, with UNET showing slightly better performance.