Szczegóły publikacji
Opis bibliograficzny
Modified U-Net with attention gate for enhanced automated brain tumor segmentation / Shoffan SAIFULLAH, Rafał DREŻEWSKI, Anton Yudhana, Maciej WIELGOSZ, Wahyu Caesarendra // Neural Computing & Applications ; ISSN 0941-0643. — 2025 — vol. 37 iss. 7, s. 5521–5558. — Bibliogr. s. 5554-5558, Abstr. — Publikacja dostępna online od: 2025-01-02. — S. Saifullah - dod. afiliacja: Universitas Pembangunan Nasional Veteran Yogyakarta, Indonesia
Autorzy (5)
- AGHSaifullah Shoffan
- AGHDreżewski Rafał
- Yudhana Anton
- AGHWielgosz Maciej
- Caesarendra Wahyu
Słowa kluczowe
Dane bibliometryczne
| ID BaDAP | 158471 |
|---|---|
| Data dodania do BaDAP | 2025-03-29 |
| Tekst źródłowy | URL |
| DOI | 10.1007/s00521-024-10919-3 |
| Rok publikacji | 2025 |
| Typ publikacji | artykuł w czasopiśmie |
| Otwarty dostęp |  |
| Czasopismo/seria | Neural Computing & Applications |
Abstract
This study addresses the formidable challenges encountered in automated brain tumor segmentation, including the complexities of irregular shapes, ambiguous boundaries, and intensity variations across MRI modalities. Manual segmentation, plagued by subjectivity and time constraints, further exacerbates the problem. To address these issues, we propose a modified U-Net architecture with an integrated attention gate. The proposed model demonstrates high performance, with notable Dice Similarity Coefficient (DSC) and Jaccard Index (JI) values across various tumor classes, consistently exceeding 0.93 and 0.87, respectively. Incorporating Contrast-Limited Adaptive Histogram Equalization and Histogram Equalization improves segmentation accuracy, particularly in cases of Meningioma. Comparative analyses against established models reveal a DSC of 0.9521 and a JI of 0.9093, underscoring the superiority of our method. Validation in the BraTS 2021 dataset underscores the robustness of the method, achieving high DSC and JI scores in four MRI modalities, with the T2 modality demonstrating the highest performance (DSC: 0.9216, JI: 0.8556). While acknowledging these achievements, we recognize challenges related to dataset specificity and computational intensity associated with the attention gate. Future research efforts should address these issues to improve the generalizability and applicability of the method in real-world scenarios. In addition to presenting a novel automated brain tumor segmentation method, this study contributes comprehensive result values and comparative analyses with previous research, providing valuable insights into the evolving landscape of medical image analysis.
