Author: Yoonha Eo 👨🔬
Affiliation: Yonsei University 🌍
Purpose: To develop a fully automatic and unsupervised algorithm for estimating the Exposure Index (EI) of various Regions of Interest in X-ray imaging using advanced foundation models. Traditional EI estimation methods, which rely on fixed masks or pixel thresholding, often fail to accurately isolate anatomical ROIs, resulting in less reliable dose calculations. This study aims to create an EI estimation framework that leverages advanced neural network architectures to segment ROIs efficiently, ensuring robust performance in real-time clinical applications.
Methods: The proposed method comprises two primary stages: Masking and Calculating. In masking stage, the PubMedCLIP model, trained on a radiology dataset, generates text-based prediction of possible anatomical regions from a predefined list. This text prediction guide YOLO-World, which detects ROIs in the X-ray images and outputs bounding boxes and point coordinates. These outputs are then used as prompts by MobileSAM to create segmentation masks, effectively isolating relevant anatomical structures from non-ROI. In calculating stage, the segmented ROI is further refined by applying Otsu’s thresholding method to binarize the image, generating a binary mask that is applied to the raw X-ray image. This allows the calculation of the average pixel intensity within the masked region, which serves as the basis for determining the Exposure Index.
Results: The proposed algorithm demonstrated high precision with real-time calculation, ensuring its suitability for integration into clinical workflows. It also exhibited strong generalizability and robustness across diverse imaging devices, patient anatomies, and exposure conditions, consistently delivering accurate pixel intensity values for the target regions.
Conclusion: The proposed algorithm demonstrates precision, efficiency, and generalizability, automatically segmenting anatomical ROIs and calculating reliable exposure indices. This method offers a standardized approach to improve X-ray image quality and reduce risks of over- and under-exposure, enhancing patient safety and diagnostic accuracy in clinical workflows.