Brain-Tumor-Segmentation

Background

Surge in cancer cases globally have led to increase in computer aided diagnosis and research in biomedical imaging and diagnostic radiology. The task of manual segmentation is rigorous, time-consuming and accurate tumor segmentation depends on the expertise of the pathologist, incorrect segmentation will cause in recurrent brain cancer or long lasting side-effects.

For accessing the interactive notebooks:

Table of Contents

1. Introduction
2. Network Architecture
3. Loss Function
4. Segmentation Results
   • Quantitative Results
   • Visualization Results
5. Conclusion

Introduction

Automatic brain tumor segmentation is one such task which will assist, improve doctors and radiologists accuracy in detecting and delineating the tumor sub-type. Automated brain tumor segmentation is highly desirable as it will help doctors learn about the prognostic factors and monitor the progression of the tumor and plan for treatment. For this we propose a method based on modified 3DUNet architecture which produced state-of-the-art segmentation results on the Brats 2020 Challenge.

Network Architecture

Loss Function

The proposed model uses Modified soft dice loss to optimize the loss function. Individual dice loss for the three labels (Enhancing tumor, Whole tumor, and Tumor core) are computed and added for the final dice loss, and the optimization is performed on the individual dice loss of the labels.

In the above equations, indicates the 3x144x144x128 matrix of the ground truth annotation and indicates the 3x144x144x128 matrix of the predicted segmentation output by the network and indicates the smoothing factor which set to 1

Segmentation Results

Quantitative Results

We calculate the proposed methods results using the statistical parameters-Dice Coefficient, Sensitivity, Specificity and Hausdorff Distance for Enhancing tumor, Whole tumor and Tumor core for the validation set. Team - Zillella Brats 2020 Challenge Leaderboard

Parameters	Dice_ET	Dice_WT	Dice_TC	Sensitivity_ET	Sensitivity_WT	Sensitivity_TC	Specificity_ET	Specificity_WT	Specificity_TC	Hausdorff95_ET	Hausdorff95_WT	Hausdorff95_TC
Mean	0.80661	0.89414	0.85721	0.81519	0.9251	0.84795	0.99971	0.99884	0.99953	23.1904	5.56554	5.44695
StdDev	0.23953	0.08177	0.12749	0.25155	0.07596	0.16678	0.00043	0.00116	0.0007	85.62448	10.87704	9.90171
Median	0.87871	0.91993	0.91058	0.89553	0.94956	0.90561	0.99985	0.99918	0.9998	1.73205	3.16228	3
25quantile	0.80214	0.87042	0.82128	0.82766	0.90596	0.81328	0.99957	0.99858	0.9995	1	2.23607	1.73205
75quantile	0.92648	0.94299	0.93417	0.94811	0.97373	0.95586	0.99996	0.99958	0.9999	3	5.47723	5.47723

Visualization Results

Axial View

                                   Brain MRI                                    Ground Truth                             Predicted Tumor

Coronal View

                                   Brain MRI                                    Ground Truth                             Predicted Tumor

Sagittal View

                                   Brain MRI                                    Ground Truth                             Predicted Tumor

                      Region in red indicates the tumor core Region in green indicates the whole tumor                                                                         Region in blue indicates the enhancing tumor region

Conclusion

With the proposed modified 3DUNet architecture state-of-the-art segmentation results for Enhancing tumor were obtained with approximatelty 1% improvement in dice coefficient when compared to the winners of Brats 2020 Challenge. The model is computationally very efficient taking less than 2 seconds to segment brain tumors from the whole brain MRI scan.