Skip to content

ksatrajit0/medbrain-net

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

16 Commits
 
 
 
 
 
 
 
 

Repository files navigation

🧠 Medbrain-net: a dual-pathway framework for brain tumor segmentation and classification from multi-modal MR images

Python TensorFlow PyTorch License: GPL v3 Kaggle

📌 Project Overview

MedBrain-Net is an advanced, automated medical imaging framework designed to assist radiologists in the initial diagnosis and treatment planning of brain tumors. By merging two specialized deep learning pathways, the system provides a comprehensive clinical profile of the tumor:

  • Precise Segmentation: High-fidelity boundary delineation using attention-gated architectures.
  • Advanced Classification: Intelligent subtype identification (Glioma, Meningioma, Pituitary).

Clinical Impact: This framework effectively minimizes inter-observer variation and accelerates diagnostic workflows, delivering stable and interpretable results in high-pressure clinical settings.


🏗️ System Architecture

MedBrain-Net operates via a dual-branch logic to ensure both spatial accuracy and categorical precision:

image

1. 🟥 Segmentation Branch (PyTorch)

An ensemble of U-Net++ models optimized for the BraTS 2020 dataset.

  • Nested Skip Connections: Captures multi-scale features for complex tumor geometries.
  • CBAM (Convolutional Block Attention Module): Dynamically suppresses noise and enhances tumor-relevant feature maps.
  • Deep Supervision: Ensures gradient flow and feature consistency across all decoding levels.
  • Hybrid Loss: Optimized via $Loss = \text{BCE} + \text{Dice} + \text{Focal Loss}$ to handle severe class imbalance.

2. 🟦 Classification Branch (TensorFlow/Keras)

A DenseNet201 backbone utilizing transfer learning on the SARTAJ dataset.

  • Feature Re-use: Dense blocks ensure maximum information flow for subtype differentiation.
  • Robust Regularization: Implements a 0.5 Dropout rate to prevent overfitting on specific scan orientations.
  • On-the-fly Augmentation: Integrated Keras layers for rotation, zoom, and horizontal flipping.

📊 Performance Benchmarks

Metric Segmentation (BraTS 2020) Classification (SARTAJ)
Accuracy 97.25%
F1-Score - 95%
Dice Coefficient 0.8237
Mean IoU 0.7051

About

Medbrain-net: a dual-pathway framework for brain tumor segmentation and classification from multi-modal MR images

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors

Languages