A Dual-Stage Deep Learning Framework for Obesity Assessment with Ultrasound Image Segmentation and Severity Classification

Obesity is a major global health challenge associated with metabolic dysfunction and multi-organ impairment. Early detection of fat accumulation is critical but depended on the resolution and operator dependency of conventional ultrasound techniques. This study proposes a two-stage deep learning framework for automated segmentation and severity classification of obesity using abdominal ultrasound images. The pipeline integrates four encoder–decoder architectures (ENet, SegNet, U-Net, and U-Net++) for precise fat-layer segmentation, followed by convolutional neural networks (LeNet-5, VGG19, ResNet50, and EfficientNetB0) for four-level obesity severity classification. A curated dataset of 550 radiologist-annotated ultrasound images was used for training and evaluation. The U-Net++ model achieved the highest segmentation performance with a Dice coefficient of 0.986 and an Intersection over Union of 0.973, while EfficientNetB0 recorded a superior classification accuracy of approximately 97% with stable validation performance. The results demonstrate that ultrasound-based deep learning pipelines can deliver reliable, non-invasive, and cost-effective solutions for early obesity detection, offering strong potential for integration into routine diagnostic practice.