Manuscript received April 16, 2024; revised July 4, 2024; accepted August 2, 2024; published February 14, 2025.

Abstract—In this paper, we have proposed a deformable encoder-decoder neural network for liver segmentation from multi-modality Computed Tomography (CT) images. Liver segmentation is a predominant step to taking conclusive action toward liver disease detection, therapeutic decision planning, and post-operation assessment. The computed tomography scan has become the default choice of medical practitioners to determine hepatic anomalies. However, due to improvements in image acquisition protocols, imaging data is growing making the manual delineation process burdensome and tedious for clinicians and becoming reliant on expert proficiency and experience. Furthermore, automatic liver segmentation is challenging due to complicated anatomy, shape variance, and less contrast variation within itself and its tumors, between its neighboring organs like the heart, and spleen, and even discontinuity in liver contours. Moreover, normal convolutions with fixed feature patterns cannot predict irregular liver patterns Thus, our proposed Def-UNet for liver segmentation is developed by modifying the encoder convolution method by deformable convolutions and skip connections by local feature recalibration which sends high-level feature information to the decoder side. The deformable convolution is computationally less expensive and best suited for shapevariant medical images. Further, the adaptive recalibration through a Squeeze-and-Refine network helps to learn the channel-wise interdependencies and gather the salient details from the fusion applied high-level features. As a bridge module, we have employed an atrous pyramid pooling module to capture the spatial information from the low-level features with the help of dissimilar receptive fields. These methods help the Def-UNet to enhance the accuracy and greatly reduce the computational burden of the other DLbased segmentation methods. The efficacy of the proposed method is experimented on two datasets Combined (CTMRI) Healthy Abdominal Organ Segmentation (CHAOS) and 3DIRCADb that are publicly available. The experimental result analysis illustrates that the proposed model has attained a dice similarity coefficient of 0.966 and 0.972 for liver segmentation.

Keywords—liver segmentation, deformable convolution, deep learning, squeeze and refine networks, encoder-decoder architecture

Cite: Bindu Madhavi A. Tummala and Soubhagya Sankar B. Barpanda, "Def-UNet with Feature Fusion and Recalibration for Liver Segmentation in Multi-Modality CT," Journal of Image and Graphics, Vol. 13, No. 1, pp. 64-72, 2025.

Copyright © 2025 by the authors. This is an open access article distributed under the Creative Commons Attribution License (CC-BY-4.0), which permits use, distribution and reproduction in any medium, provided that the article is properly cited, the use is non-commercial and no modifications or adaptations are made.