Comprehensive benchmarking of CNN-based tumor segmentation methods using multimodal MRI data.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Magnetic Resonance Imaging (MRI) is vital for brain tumor detection.
• Manual tumor segmentation from MRI is time-consuming and labor-intensive.
• Automated segmentation methods are increasingly important for efficient and accurate analysis.

Purpose of the Study:

• To benchmark and evaluate four Convolutional Neural Network (CNN)-based methods for brain tumor segmentation.
• To compare the performance of CaPTk, 2DVNet, EnsembleUNets, and ResNet50.
• To assess segmentation accuracy using direct image comparison and radiomic features.

Main Methods:

• Utilized 1251 multimodal MRI scans from the BraTS2021 dataset.
• Compared four CNN models: CaPTk, 2DVNet, EnsembleUNets, and ResNet50.
• Evaluated performance using Dice Similarity Coefficient (DSC) and Hausdorff Distance (HD), and radiomic features (CCC, TDI, RMSE).

Main Results:

• EnsembleUNets achieved the highest performance with a DSC of 0.93 and HD of 18 on the BraTS2021 dataset.
• Radiomic feature analysis confirmed EnsembleUNets' superior precision (CCC=0.79, TDI=1.14, RMSE=0.53).
• Validation on the UPENN-GBM dataset showed EnsembleUNets maintained high accuracy (DSC=0.85, HD=17.5).

Conclusions:

• EnsembleUNets significantly outperforms other evaluated CNN methods for brain tumor segmentation in MRI.
• The findings support the use of EnsembleUNets for accurate and efficient brain tumor segmentation.
• This research aids in informed decision-making for improved diagnosis, treatment, and prognosis of brain tumors.