MNR2NeXt-50: Segmentation and quantification of epicardial fat from cardiac CT images using transfer learning with an optimized ensemble model

Abstract

Excessive epicardial adipose tissue (EAT) presents a significant risk for cardiac issues. The level of EAT can be assessed using CT images of the heart. Machine learning and deep learning techniques are available to quantify EAT through the segmentation and analysis of CT images automatically. These techniques require precise segmentation and faster processing. This study proposes a hybrid method that combines the ResNeXt-50 deep learning model with Naïve Bayes-based ensemble learning and Manta Ray Foraging Optimization (MRFO) to enhance the segmentation process. Various Naïve Bayes ensemble configurations, such as Random Forest, AdaBoost, XGBoost, Extra Trees, CatBoost, and Generalized Linear Model were analysed to identify the optimal configuration. The techniques were evaluated on images with resolutions of 128 × 128, 256 × 256, and 512 × 512 pixels using metrics including accuracy, precision, recall, Dice score, and processing time. These models were trained and validated with a dataset having 878 CT slices from 20 patient scans, utilizing an 80-20 train-test split and five-fold cross-validation for robustness. Among the tested configurations, the Naïve Bayes with Random Forest ensemble (MNR²NeXt-50) demonstrated better performance, yielding an accuracy of 98.88 %, a Dice score of 98.79 %, and a processing time of 1.25 s per image for 512 × 512 pixels at 5-fold cross-validation. Further, compared to other existing methods, the MNR²NeXt-50 shows improvements in accuracy (0.12 %–12.88 %), Dice score (0.06 %–31.59 %), and processing time (1.152–1504 times faster). Future efforts will focus on clinical validation, expanding the dataset through multi-center collaborations, exploring integration with multi-modal imaging, and investigating the association between EAT and other clinical indicators to facilitate broader clinical adoption.