Tversky Loss Mechanisms: A ResUNet Approach to Improving Brain Tumor Segmentation.
Date:
In this talk, we explore the implementation of our research titled “Tversky Loss Mechanisms: A ResUNet Approach to Improving Brain Tumor Segmentation.”
Introduction
Biomedical imaging plays a critical role in healthcare, particularly in detecting and diagnosing brain tumors. Brain tumors—whether benign or malignant—can significantly disrupt brain function. Early detection is especially important because malignant tumors can grow rapidly.
Magnetic Resonance Imaging (MRI) is widely used for brain imaging. However, manually segmenting tumors from MRI scans is labor-intensive, time-consuming, and often inconsistent due to differences between observers. To address these challenges, our work proposes a deep-learning-based automated segmentation approach.
Objectives
The main objectives of this research are:
- To improve brain tumor segmentation accuracy using a ResUNET model.
- To address class imbalance using the Tversky loss function, which helps detect small tumor regions.
- To improve segmentation performance, especially for small and critical tumor areas.
- To evaluate the proposed model and compare it with traditional CNN and U-Net methods.
Methodology
A dual-model approach was adopted:
- First, a ResNet-50 classifier, pre-trained on ImageNet, was used to detect whether a tumor is present.
- If a tumor was detected, the image was passed to a ResUNET model for segmentation.
The ResUNET architecture integrates:
- Residual connections
- Convolutional layers
- Max pooling and ReLU activation in the encoder
- A decoder to reconstruct spatial features and generate a binary segmentation mask
To handle class imbalance, the Tversky loss function was used during training, improving sensitivity to small tumor pixels.
Training used:
- Generator-based approach
- Batch size of 16
- Early stopping to prevent overfitting
Performance was evaluated using standard segmentation metrics.
Metrics and Evaluation
To evaluate performance, the following metrics were used:
- Dice Coefficient
- Intersection over Union (IoU)
- Precision, Recall, and F1-score
Results
- Dice Score: 0.98
- IoU: 0.95
- Classification Accuracy: 98%
These results indicate strong agreement between predicted masks and ground truth annotations.
Comparative Study
The model was compared with several recent approaches, including:
- SEResU-Net
- ResUNet+
- U-Net++
- CNN + U-Net Ensemble
The proposed model achieved higher Dice and IoU scores than many existing methods, demonstrating improved segmentation performance.
Results Visualization
The study includes:
- Segmentation outputs comparing original images, ground-truth masks, and predicted masks
- Training and validation accuracy curves
- Loss curves
- Confusion matrix results
These visual results confirm the robustness of the proposed method.
Conclusion
The proposed ResUNET model with Tversky loss provides a robust and efficient solution for brain tumor segmentation.
Key strengths:
- High segmentation accuracy
- Effective handling of class imbalance
- Improved detection of small tumor regions
- Efficient dual-stage architecture
This approach can support faster, more consistent diagnosis and improved treatment planning.
Future Work
Future directions include:
- Cross-validation and testing on external datasets
- Improving generalization
- Extending the model to 3D MRI scans
- Applying the framework to other cancer types
Acknowledgment
We would like to thank all collaborators and institutions who supported this research and contributed to its successful completion.
Feel free to contribute, raise issues, or suggest improvements!
After Technical Session
