Combining DeepLabV3 with Attention Mechanisms for Accurate Brain Tumor Segmentation: Insights from BraTS 2020 and a Private Clinical Dataset

• Main Authors: Ranjbarzadeh, Ramin, Keles, Ayse, Ozisik, Pinar Akdemir, Demirci, Harun, Kul, Halil, Crane, Martin, Anari, Shokofeh, Bendechache, Malika
• Format: Conference Proceeding
• Language: English
• Subjects: Accuracy; Atrous Spatial Pyramid Pooling; Attention Mechanism; Attention mechanisms; Brain modeling; Brain Tumor Segmentation; DeepLabV3; Image segmentation; Information processing; Measurement; Medical diagnosis; Overfitting; Urban areas; Visualization
• ISSN: 2471-8963
• DOI: 10.1109/EUVIP61797.2024.10772776

This study explores the effectiveness of the DeepLabV3 model and its variations, which include attention mechanisms, in accurately segmenting brain tumors. The study used both the publicly accessible BraTS 2020 dataset and a private clinical dataset obtained from Ankara Bilkent City Hospital, Turkiye. The research assesses the impact of attention mechanisms on segmentation performance for three tumor classes: Tumor Core, Edema, and Enhancing Tumor. These attention mechanisms are incorporated both inside the overall model and especially in the Atrous Spatial Pyramid Pooling (ASPP) block. The accuracy and dependability of the models are evaluated using key metrics such as Precision, Recall, Intersection over Union (IoU), and Dice coefficient. The findings suggest that attention mechanisms can improve the identification of features, but their impact is highly dependent on their setup and the specific dataset being used. The conventional DeepLabV3 model frequently achieves better results compared to its attention-enhanced variations, especially when dealing with difficult tumor types and datasets. This indicates the possibility of overfitting or misalignment in the priority of features. This study enhances our comprehension of how attention mechanisms can be effectively included in medical imaging models, emphasizing the importance of meticulous implementation and calibration to optimize their potential. The results offer crucial insights for the development of more precise and robust segmentation algorithms in medical diagnostics.