Model Name: NeuroNest3D

Model Description

Overview

Model Architecture

• Multiple convolutional layers with ReLU activation and 'same' padding for feature learning.
• MaxPooling layers for spatial down-sampling.
• UpSampling layers coupled with Conv2D layers for expanding the feature maps.
• Concatenation of feature maps from different stages to retain spatial information.
• A final Conv2D layer with softmax activation for multi-class segmentation.

Training

• Dataset: Trained on the BraTS 2020 dataset, comprising diverse brain MRI scans with annotations.
• Preprocessing and Data Generation: Implemented using a custom DataGenerator class for efficient batch processing and augmentation.
• Loss Function: Utilizes a categorical cross-entropy loss function.
• Metrics: Includes accuracy, Mean IoU, and custom-defined dice coefficient metrics for each tumor type (necrotic, edema, enhancing).
• Optimization: Employs the Adam optimizer with a learning rate of 0.001.
• Callbacks: Early stopping, reduce learning rate on plateau, model checkpointing, and CSV logging for monitoring the training process.

Performance Metrics

• Overall Dice Coefficient
• Per-class Dice Coefficients (Necrotic, Edema, Enhancing)
• Precision, Sensitivity, Specificity

How to Load The Model