Demystifying deep learning-based brain tumor segmentation with three-dimensional UNets and explainable artificial intelligence: A comparative analysis

Ming Jie Ong1, Sze Yinn Ung2, Sim Kuan Goh1, *, Jimmy Y. Zhong3, *

1School of Artificial Intelligence and Robotics, Xiamen University Malaysia, Sepang, Selangor, Malaysia;

2School of Computing and Data Science, Xiamen University Malaysia, Sepang,

Selangor, Malaysia;

3Georgia State/Georgia Tech Center for Advanced Brain Imaging, Georgia Institute of Technology, Atlanta, GA, USA

*Correspondence to: Sim Kuan Goh, PhD, simkuan.goh@xmu.edu.my; Jimmy Y. Zhong, PhD, jzhong34@gatech.edu.

Funding: The study was supported by the Ministry of Higher Education Malaysia through the Fundamental Research Grant Scheme (No. FRGS/1/2023/ICT02/XMU/02/1) and Xiamen University Malaysia through the Xiamen University Malaysia Research Fund (Nos. XMUMRF/2022-C10/IECE/0039 and

XMUMRF/2024-C13/IECE/0049; both to SKG).

Objective: Accurate brain tumor segmentation from magnetic resonance imaging is essential but labor intensive and variable across clinicians. Deep learning models, especially three-dimensional U-shaped artificial neural network (UNet)-based architectures trained on datasets such as BraTS 2020, aim to automate this process while improving reliability and interpretability. The current study investigated the use of explainable artificial intelligence to improve the accuracy of brain tumor segmentation in magnetic resonance imaging images, with the goal of assisting physicians in clinical decision-making. This study focused on the application of UNet models for brain tumor segmentation and the use of explainable artificial intelligence techniques such as gradient-weighted class activation mapping (Grad-CAM) and attention-based visualization to enhance the understanding of these models.

Methods: Three deep learning models, UNet, residual UNet (ResUNet), and attention UNet (AttUNet), were evaluated to identify the best-performing model. Explainable artificial intelligence was employed with the aims of clarifying model decisions and increasing physicians’ trust in these models. We compared the performance of two UNet variants (ResUNet and AttUNet) with that of conventional UNet in segmenting brain tumors from the BraTS 2020 public dataset and analyzed model predictions with Grad-CAM and attention-based visualization. Using the latest computer hardware, we trained and validated each model using the Adam optimizer and assessed their performance with respect to (i) training, validation, and inference times; (ii) segmentation similarity coefficients and loss functions; and (iii) classification performance.

Results: Notably, during the final testing phase, ResUNet outperformed the other models with respect to Dice and Jaccard similarity scores, as well as accuracy,

recall, and F1 scores. Grad-CAM provided visuospatial insights into the tumor subregions that each UNet model focused on, while attention-based visualization

provided valuable insights into the working mechanisms of AttUNet’s attention modules.

Conclusion: These results indicate that ResUNet is the best-performing model, and we conclude by recommending its use for automated brain tumor segmentation in future clinical assessments.

Key Words: attention mechanism; brain tumor segmentation; BraTS 2020; Dice score; explainable AI; explainable attention; Grad-CAM; residual network; UNet comparison; UNet