Manuscript received October 1, 2025; revised October 17, 2025; accepted January 7, 2026; published May 27, 2026.

Abstract—Epicardial Adipose Tissue (EAT) volume has been reported as a clinically relevant biomarker of metabolic risk, yet conventional diabetes assessment relies heavily on biochemical measurements that fluctuate with short-term physiological changes. Computed Tomography (CT)–based EAT quantification offers a more stable alternative; however, accurately delineating the epicardial boundary remains difficult due to its thin, low-contrast structure. Existing Convolutional Neural Network (CNN) approaches trained with Dice-based losses often overlook subtle boundary regions, leading to fragmented contours and degraded downstream EAT estimation. To address this limitation, we propose a Hybrid Focal Loss (HFL) that adaptively balances focal and regional penalties to enhance learning of fine epicardial structures. The loss is embedded in a two-stage framework: Stage 1 performs epicardium segmentation, and Stage 2 reconstructs a smooth mask via spherical harmonics interpolation to enforce global surface consistency. Experimental results on CT images from 15 patients demonstrate that HFL improves segmentation performance by 1.6% in F1-Score compared to Dice Binary Cross-Entropy (BCE) loss and further enhances reconstructed mask accuracy, reducing Hausdorff distance by 6.0 points. When applied to EAT extraction, HFL yields the highest F1 and Intersection over Union (IoU) scores across all evaluated losses, translating into improved agreement with the ground truth of EAT volumes. These findings demonstrate that HFL strengthens both boundary sensitivity and volumetric reliability, suggesting its value for robust EAT quantification and cardiometabolic risk I. INTRODUCTION Diabetes mellitus is a chronic metabolic disorder characterized by sustained hyperglycemia resulting from impaired insulin secretion or sensitivity. Insulin, secreted by pancreatic β-cells, plays a central role in maintaining glucose homeostasis; its dysfunction leads to a range of systemic complications including retinopathy, nephropathy, and arteriosclerosis. Diabetes is typically classified into two major types: Type 1, caused by autoimmune destruction of β-cells, and Type 2, which is predominantly associated with genetic predisposition and lifestyle factors such as diet and physical inactivity. Although clinical diagnosis is conventionally performed using fasting plasma glucose or glycated hemoglobin (HbA1c) levels, these biochemical indicators are affected by short-term physiological fluctuations and comorbidities, thereby reducing diagnostic robustness. Consequently, the identification of more stable, imaging-based biomarkers is essential for reliable metabolic assessment. assessment.

Keywords—diabetes, epicardial adipose tissue, epicardium, Computed Tomography (CT) images, segmentation, hybrid focal loss, spherical harmonics

Cite: Ibuki Naka, Yinhao Li, Yutaro Iwamoto, Jain Rahul Kumar, Xianhua Han, Atsuyuki Wada, Yuji Tezuka, Kiyosumi Maeda, Atsunori Kashiwagi, and Yen-Wei Chen, "Hybrid Focal Loss for Accurate Epicardium Segmentation and Mask Generation via Spherical Harmonics," Journal of Image and Graphics, Vol. 14, No. 3, pp. 403-411, 2026.

Copyright © 2026 by the authors. This is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited (CC BY 4.0).