Brain Tumor Segmentation

Model Description

This model is a U-Net based deep learning model for the segmentation of brain tumors from 2D MRI slices. It is an end-to-end segmentation model designed to assist radiologists by automatically identifying and outlining tumorous regions. The model leverages transfer learning by using a pre-trained ResNet50 backbone as its encoder.

Intended Use & Limitations

This model is intended for research and educational purposes. It is a tool for exploring deep learning for medical image segmentation.

Limitations:

The model was trained on a specific dataset and may not generalize well to other datasets from different scanners, protocols, or patient populations.
It should not be used for clinical diagnosis without validation by a qualified medical professional.

Training Data

The model was trained on the LGG MRI Segmentation Dataset from Kaggle.

Volume: 3,929 MRI images from 110 patients.
Annotation: Expert-annotated binary masks for tumor regions.
Image Size: 256x256 pixels with 3 channels (converted to RGB).

Training Procedure

The model was trained using a two-stage fine-tuning approach with PyTorch.

Model Architecture:

Architecture: U-Net
Encoder: ResNet50 with imagenet pre-trained weights.
Decoder: A symmetric expanding path with skip connections.
Output: A single-channel probability map.

Hyperparameters:

Optimizer: AdamW
Loss Function: A composite loss function combining Dice Loss and Binary Cross-Entropy (BCE).

$L_{\text{combo}} = L_{\text{dice}} + L_{\text{BCE}}$

Epochs:
- Stage 1 (Frozen Encoder): 5 epochs at a learning rate of 1e−4.
- Stage 2 (Fine-tuning): 15 epochs at a learning rate of 1e−5.
Learning Rate Scheduler: ReduceLROnPlateau with patience=3, factor=0.1.
Batch Size: 16

Evaluation Results

The model was evaluated on an unseen test set composed of 10% of the patient data.

Metric	Value
Dice Score	0.8898
IoU Score	0.8015
Accuracy	0.9963
Precision	0.8820
Recall	0.8978