Boosting Brain Tumor Detection Accuracy in MRI Using Transfer Learning and Fine-Tuned DeepLabv3+

Ahmed Soomro, Toufique, Fatihi, Ayoub, Zheng, Lihong, Awais, Muhammad ORCID: https://orcid.org/0000-0001-6421-9245, Jamal Afifi, Ahmed and Arnaout, Jean-Paul (2026) Boosting Brain Tumor Detection Accuracy in MRI Using Transfer Learning and Fine-Tuned DeepLabv3+. IEEE Open Journal of the Communications Society, 7. pp. 1038-1050. ISSN 2644-125X

Preview

PDF (Boosting_Brain_Tumor_Detection_Accuracy_in_MRI_Using_Transfer_Learning_and_Fine-Tuned_DeepLabv3) - Published Version Available under License Creative Commons Attribution. Download (2MB) | Preview

Abstract

Detecting and diagnosing brain tumors in adults is a complex task, given their varied appearance and serious health risks. Manual analysis of magnetic resonance imaging (MRI) scan data is time-consuming and error-prone, affecting the final analysis and treatment. In this work, we implemented a deep learning method based on the convolutional neural network (CNN) framework using the DeepLabv3+ model to segment brain tumor types, including glioma, meningioma, and pituitary tumors, from the contrast-enhanced magnetic resonance imaging (CE-MRI) database. The method is based on a ResNet-34 backbone, Atrous Spatial Pyramid Pooling (ASPP), and a boundary-aware decoder, all of which are specifically trained on CE-MRI tumor images. To address the imbalance between tumor and background pixels in the image, we used a Weighted Binary Cross-Entropy (WBCE) loss function, which improves segmentation of smaller, more irregular tumor regions. We experimented with every step that ensures the method's capability, including a step-by-step ablation study to examine the contribution of each stage. The method is evaluated on CE-MRI databases, both internal and external, without any additional fine-tuning, and on multiple BraTS benchmark database editions (2020, 2021, 2023). The method achieves strong performance across all databases, with a Dice similarity coefficient of 98%, a sensitivity of 99.3%, and an accuracy of 99.1%. This demonstrates the effective implementation of the method, which combines the CNN architecture DeepLabv3+ with a loss function that outperforms existing methods in validations.

Item Type:	Article
Uncontrolled Keywords:	brain tumor,deeplabv3+,convolution neural networks,fine tuning,magnetic resonance imaging (mri),transfer learning,general computer science,sdg 3 - good health and well-being ,/dk/atira/pure/subjectarea/asjc/1700/1700
Faculty \ School:	Faculty of Science > School of Computing Sciences
UEA Research Groups:	Faculty of Science > Research Groups > Data Science and AI Faculty of Science > Research Groups > Health Computing
Related URLs:	https://www.scopus.com/pages/publication...
Depositing User:	LivePure Connector
Date Deposited:	07 Oct 2026 08:23
Last Modified:	09 Oct 2026 19:46
URI:	https://ueaeprints.uea.ac.uk/id/eprint/104737
DOI:	10.1109/OJCS.2026.3688451

Downloads

Actions (login required)

View Item