NeuroLens - MRI Brain Tumor Detection & Segmentation Pipeline

---

Overview

An end-to-end deep learning solution for automated brain tumor detection and segmentation in MRI scans, developed during my research internship at DRDO-INMAS.

This project accelerates radiological workflows using 2D/3D MRI datasets and scalable AI pipelines. The pipeline includes detection (YOLOv8), 3D U-Net segmentation.

---

Architecture

---

BraTS2020 Dataset Overview

• Gold standard for brain tumor segmentation (MICCAI BraTS challenge)
• 369 multi-institutional 3D MRI volumes (T1, T1ce, T2, FLAIR)
• Expert-annotated masks: necrotic core, edema, enhancing tumor
• 240×240×155 voxels, 1mm³ isotropic, skull-stripped, co-registered
• Used for training, validation, and benchmarking of segmentation models

---

YOLOv8 for Brain Tumor Detection

While the main focus of this project is on 3D U-Net-based segmentation using the BraTS dataset, we also explored YOLOv8 for fast tumor detection on 2D MRI slices. YOLOv8 enables rapid localization of tumor regions, which can be useful for pre-screening or as a first step before detailed segmentation.

• Model: YOLOv8 (PyTorch, Ultralytics)
• Reference Dataset: Medical Image Dataset: Brain Tumor Detection (Kaggle)
• Classes: Glioma, Meningioma, Pituitary
• Pipeline: Data preparation, YOLOv8 training, evaluation (mAP, precision, recall), and batch inference on test images.
• Notebook: notebooks/yolov8-brain-tumor-detection.ipynb

Key Features:

• Fast detection of tumor types in 2D MRI slices.
• Useful for rapid triage and as a pre-processing step for 3D segmentation.
• Visualizations of detection results and training metrics included.

---

3D U-Net for Tumor Segmentation

The core of this project is a 3D U-Net model trained on the BraTS2020 dataset for voxel-wise segmentation of brain tumor subregions.

• Model: 3D U-Net (Keras/TensorFlow)
• Dataset: BraTS2020 (multi-modal 3D MRI)
• Pipeline: Preprocessing, model training, evaluation, and inference.
• Notebook: notebooks/u-net-brain-tumor-segmentation.ipynb

Key Features:

• Multi-class segmentation: necrotic core, edema, enhancing tumor.
• High Dice scores on all tumor regions.
• Streamlit app for interactive inference and visualization.

---

Predictions & Results

YOLOv8 Detection Results

Glioma	No Tumor	Meningioma	Pituitary

---

3D U-Net Segmentation Result (BraTS2020)

U-Net Segmentation

---

Key Metrics (BraTS2020 & Pipeline)

Task	Metric	Value
Detection	mAP@50	91.4%
Detection	Precision	90.8%
Detection	Recall	86.5%
Segmentation	Dice Score (mean)	0.98
Region-wise	Dice (Enhancing Tumor)	0.79
Region-wise	Dice (Tumor Core)	0.83
Region-wise	Dice (Whole Tumor)	0.87

---

Usage

All model training and experiments were conducted on Kaggle Notebooks for GPU acceleration and easy dataset access.

1. Open the notebooks directly on Kaggle:

   • 3D U-Net Segmentation:
     u-net-brain-tumor-segmentation.ipynb
   • YOLOv8 Detection:
     yolov8-brain-tumor-detection.ipynb

2. To run locally (optional):

   • Download the notebooks and datasets from Kaggle.
   • Create and activate a conda environment:

     conda create -n brats python=3.8 -y
     conda activate brats

   • Install dependencies:

     pip install -r requirements.txt

   • Launch Jupyter Notebook:

     jupyter notebook notebooks/u-net-brain-tumor-segmentation.ipynb
     jupyter notebook notebooks/yolov8-brain-tumor-detection.ipynb

Note: For best results and GPU support, use Kaggle Notebooks with the provided datasets.

---

References

• BraTS 2020 Dataset
• Medical Image Dataset: Brain Tumor Detection (YOLOv8)