Comprehensive Benchmarking of CNN-Based Tumor Segmentation Methods Using Multimodal MRI Data

Abstract Magnetic resonance imaging (MRI) is become an essential and a frontline technique in the detection of brain tumor. However, manual segmentation of tumor from MRI scans is a time-consuming and labour-intensive process. There is a prevalent trend in employing fully automated methods for accurate tumor segmentation using MRI scans. The precision in brain tumor segmentation is essential for the better diagnosis, treatment and prognosis. This study focuses on benchmarking and evaluating the performance of four widely used convolutional neural network (CNN) based brain tumor segmentation methods CaPTk, 2DVNet, EnsembleUNets, and ResNet50. We used 1251 multimodal MRI scans from the BraTS2021 dataset, which encompasses T1, T2, T1ce, and Flair imaging modalities. We compared the performance of these CNN-based methods against a reference set of previously segmented images obtained with the assistance of radiologists. The evaluation encompasses direct utilization of segmented outputs and also by employing radiomic features. Performance evaluation with direct segments method using Dice Similarity Coefficient score (DSC) and Hausdorff Distance (HD) suggested better performance of EnsembleUNets with DSC and HD of 0.93 and 18 respectively, outperforming the other methods. Comparative analysis using radiomics features also revealed that EnsembleUNets is the most precise segmentation method as compared to CaPTk, 2DVNet, and ResNet50. EnsembleUNets achieved Concordance Correlation Coefficient (CCC), Total Deviation Index (TDI), and Root Mean Square Error (RMSE) of 0.79, 1.14 and 0.53 respectively and outperformed its counterparts. These findings contribute valuable insight into the comparative efficacy of EnsembleUNets, facilitating informed decision for accurate brain tumor segmentation.