Title: ADB-Net: an attention-based dilated bridge model for fully automatic intra-tumour segmentation of gliomas

Authors: Radhika Malhotra; Barjinder Singh Saini; Savita Gupta

Addresses: Department of Electronics and Communication, Dr. B.R. Ambedkar National Institute of Technology, Jalandhar, Punjab, 144011, India ' Department of Electronics and Communication, Dr. B.R. Ambedkar National Institute of Technology, Jalandhar, Punjab, 144011, India ' Department of Computer Science and Engineering, UIET, Sector 25 Panjab University, Chandigarh, 160023, India

Abstract: Glioma segmentation is a complicated task due to the non-uniform and unstructured morphology of gliomas. Moreover, the requirement for trainable parameters grows exponentially with architectural advancements. In this work, a lightweight and modified attention-based dilated bridge net (ADB-Net) architecture is developed for accurate segmentation of glioma sub-regions. The proposed work has four main benefactions. Firstly, the bridging network in D-Link is enhanced by incorporating a deformed residual connection after each dilation convolutional block to promote the mapping of multi-level information between encoding/decoding units. Secondly, a proper selection of the dilation factor is included for dilated convolutional blocks. Thirdly, four modified attention skip modules (ASM) are also introduced to provide recognition of varied-sized tumours. Lastly, the proposed architecture outperforms its baselines while minimising the number of trainable parameters by more than 50%. It achieves dice scores for the complete tumour, tumour core, and enhancing tumour as 0.971, 0.979, and 0.962, respectively.

Keywords: D-Link; attention; glioma; segmentation; loss function; BraTS; convolutional neural networks; CNN; attention skip modules; ASM.

DOI: 10.1504/IJBET.2024.137346

International Journal of Biomedical Engineering and Technology, 2024 Vol.44 No.3, pp.271 - 287

Received: 28 Feb 2023
Accepted: 07 May 2023
Published online: 13 Mar 2024 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article