Brain Tumor Segmentation using nnU-Net

Project Domain / Category

Deep Learning based Web Application

Abstract / Introduction

Brain tumor segmentation is a critical task in medical image analysis, enabling precise detection of tumor regions for diagnosis and treatment planning. In this project, you will develop a deep learning-based segmentation model using nnU-Net. nnU-Net is a model that belongs to U-Net family which is used to automatically segment brain tumors from MRI scans in the BraTS 2021 dataset.

In addition to training and evaluating the segmentation model, you will develop a Flask-based web application that allows users to upload MRI images and visualize the predicted segmentation results alongside ground truth masks.

Functional Requirements:

The following are basic requirements:

        Implement nnU-Net for brain tumor segmentation using PyTorch.

        Preprocess the BraTS 2021 dataset (multi-modal MRI images).

        Train the model using Dice Loss + Cross-Entropy Loss and evaluate its performance.

        Display segmentation results as images with ground truth comparison.

        Develop a Flask web application where users can upload an MRI scan and receive the predicted segmentation mask.

Note:

You can take introduction of the nnU-Net from the following link:

https://sh-tsang.medium.com/brief-review-nnu-net-a-self-configuring-method-for-deep-learning-based-biomedical-image-97fedf4b2079

For your better understanding the guidelines that how to start and study/code the details of all

above points is as follows:

Task 1: Understanding the Dataset

Study the U-Net model and nnU-Net model.

        Study the BraTS 2021 dataset (FLAIR, T1, T1ce, T2 modalities).

        Understand tumor segmentation labels (Whole Tumor, Tumor Core, Enhancing Tumor).

        Code to visualize sample MRI images and corresponding masks.

Task 2: Preprocessing & Data Augmentation

        Load and preprocess MRI scans using Nibabel.

        Normalize images and apply augmentations (flipping, rotation, resizing).

        Convert segmentation masks into one-hot encoded format.

Task 3: Implementing nnU-Net for Brain Tumor Segmentation

        Implement the nnU-Net model using PyTorch.

        Define loss function (Dice Loss + Cross-Entropy) and optimizer (AdamW).

        Train the model and monitor performance metrics.

Page 36 of 167