Abstract
Abdominal computed tomography (CT) scans are frequently performed in clinical settings. Opportunistic CT involves repurposing routine CT images to extract diagnostic information and is an emerging tool for detecting underdiagnosed conditions such as sarcopenia, hepatic steatosis, and ascites. This study utilizes deep learning methods to promote accurate diagnosis and clinical documentation. We analyze 2,674 inpatient CT scans to identify discrepancies between imaging phenotypes (characteristics derived from opportunistic CT scans) and their corresponding documentation in radiology reports and ICD coding. Through our analysis, we find that only 0.5%, 3.2%, and 30.7% of scans diagnosed with sarcopenia, hepatic steatosis, and ascites (respectively) through either opportunistic imaging or radiology reports were ICD-coded. Our findings demonstrate opportunistic CT's potential to enhance diagnostic precision and accuracy of risk adjustment models, offering advancements in precision medicine.
Introduction
Every year, millions of abdominal CT scans are performed to address specific clinical indications. However, these images contain significant "opportunistic" data regarding a patient’s body composition—including muscle mass, liver fat, and fluid accumulation—that often remains unquantified due to the time constraints of manual radiological assessment. Conditions such as sarcopenia (muscle wasting), hepatic steatosis (fatty liver), and ascites (abdominal fluid) are critical predictors of mortality and clinical complications, yet they are frequently underdiagnosed.
This research addresses the "coding gap": the discrepancy between clinical findings visible on imaging and their formal documentation as ICD codes in Electronic Health Records (EHR). By automating opportunistic imaging, the study aims to improve diagnostic precision and risk adjustment without necessitating additional diagnostic procedures.
Methods
The researchers developed a multi-stage deep learning pipeline to analyze 2,674 inpatient CT scans from a single academic medical center (2014–2018). The pipeline integrates several key components:
- Tissue Segmentation (Comp2Comp): An open-source library utilizing Convolutional Neural Networks (CNNs) to segment tissues at the third lumbar vertebra (L3) level, enabling the calculation of skeletal muscle area and liver/spleen density.
- Ascites Detection (Merlin): A 3D vision-language foundation model fine-tuned on 200 expert-labeled scans. Merlin leverages large-scale pre-training to detect fluid accumulation with high reliability.
- Automated Labeling (STARR-Labeler): This tool was used to extract existing ICD-9/10 codes and radiology report text to establish a baseline for comparison.
Clinical metrics, such as Hounsfield Units (HU) for liver density and T-scores or BMI-adjusted Z-scores for muscle mass, were used to define the imaging-based "ground truth."
Results
The analysis revealed a profound disconnect between imaging evidence and official clinical documentation across all three conditions studied.
- Sarcopenia: Opportunistic CT identified significant muscle loss in 14.6% to 19.3% of the cohort. Despite this high prevalence on imaging, only 0.5% of these patients received an official ICD code for sarcopenia.
- Hepatic Steatosis: Only 3.2% of scans showing fatty liver via opportunistic imaging or mentioned in radiology reports were formally coded.
- Ascites: While better recognized than the other conditions, only 30.7% of imaging-detected cases were ICD-coded. The Merlin model achieved an AUROC of 90.3% for ascites detection.
The results suggest that even when findings are mentioned in narrative radiology reports, they often fail to be translated into formal diagnostic codes, indicating a breakdown in the clinical documentation workflow.
Discussion and Conclusion
This study demonstrates that deep learning can effectively "rescue" diagnostic data from routine imaging. The success of the Merlin foundation model suggests a shift toward more flexible AI architectures capable of understanding complex relationships between medical images and clinical text.
However, several limitations remain. The single-center, retrospective nature of the study may reflect specific local coding cultures. Furthermore, the "human in the loop" remains a critical factor; in many systems, professional coders require explicit physician documentation rather than just radiology report mentions to assign a code. Future work should focus on integrating these automated findings directly into physician workflows to bridge the administrative gap and expand the application of opportunistic imaging to other conditions like osteoporosis and cardiovascular disease.
Summary updated