Related Experiment Videos
AI-Assisted cardiomegaly screening via implicit morphological inference and human-in-the-loop validation
Muhammad Masdar Mahasin1, Agus Naba1, Chomsin S Widodo1
1Department of Physics, Brawijaya University, Malang, Indonesia.
Journal of X-Ray Science and Technology
|July 17, 2026
Summary
This study introduces UBNet-Seg, a lightweight AI model for efficient cardiomegaly screening by segmenting lung fields. Human-in-the-loop refinement significantly improves accuracy, offering a robust solution for clinical settings.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Computer Vision
Background:
- Manual Cardiothoracic Ratio (CTR) measurement for cardiomegaly screening is a clinical bottleneck.
- Current deep learning models for this task can be computationally expensive and lack interpretability.
Purpose of the Study:
- To develop a resource-efficient and interpretable framework for cardiomegaly screening using implicit morphological inference.
- To bypass the need for explicit heart segmentation by using lung field segmentation as a proxy.
Main Methods:
- Developed UBNet-Seg, a lightweight U-Net variant (2.3M parameters), trained on 11,748 images to segment lung fields.
- Evaluated performance on external NIH and OpenI datasets, assessing automated accuracy and inference time.
- Integrated a Human-in-the-Loop mechanism to refine model predictions and address domain shifts.
Main Results:
- UBNet-Seg achieved a lung Dice Coefficient of 95.85% and an inference time of 0.05s.
- Fully automated accuracy reached 90.31% (NIH) and 76.07% (OpenI).
- Expert-guided refinement significantly improved accuracy to 93.63% (NIH) and 91.21% (OpenI) (p < 0.001).
Conclusions:
- Medial lung boundaries serve as reliable structural proxies for cardiomegaly screening.
- UBNet-Seg offers a robust, explainable, and low-latency alternative to manual CTR measurement.
- The framework is suitable for resource-constrained clinical settings, especially with Human-in-the-Loop integration.