Document Type : Original/Review Paper

Authors

• Omid Khalaf Beigi ¹
• Seyed Alireza Bashiri Mosavi ²

¹ Department of Electrical and Computer Engineering, Kharazmi University, Tehran, Iran

² Department of Electrical and Computer Engineering, Buein Zahra Technical University, Buein Zahra, Qazvin, Iran

Abstract

A brain tumor is one of the most serious and life-threatening brain diseases that can profoundly affect an individual’s life. Accordingly, the present study addresses the challenge of refining brain tumor segmentation based on Magnetic Resonance Imaging (MRI) data and deep reinforcement learning. Although supervised learning–based approaches have shown satisfactory performance in tumor segmentation and localization, they often suffer from high uncertainty errors along tumor boundaries. In this research, a learning framework combining a supervised model with deep reinforcement learning—referred to as DURL-Net—is proposed for segmentation and refinement purposes. Specifically, the framework first employs a U-Net architecture to generate an initial segmentation mask. This initial output and the corresponding MRI are then partitioned into localized patches, which are sequentially processed by a Deep Q-Network (DQN) agent. The DQN agent interacts with the environment by selecting optimal morphological operations (such as dilation and erosion) to refine tumor boundaries and correct uncertainties patch by patch. The dataset used in this study comprises 3,064 T1-Weighted Contrast-Enhanced MRI images, employed for both segmentation and tumor-type classification tasks. Experimental results demonstrate that DURL-Net achieved a Dice Similarity Coefficient (DSC) of 86.73%, a Jaccard Index (IoU) of 78.68%, a Kappa coefficient (Kap) of 85.21%, a Sensitivity of 87.68%, and a Specificity of 96.06%.

Keywords

• MRI
• brain tumor
• deep reinforcement learning
• Deep Q-Network
• image segmentation

Main Subjects

• H.3. Artificial Intelligence