# When the Diaphragm Deceives: Subpulmonic Effusion Flagged by AI and Confirmed by Ultrasound

## TL;DR

An AI system flagged a subtle subpulmonic effusion on a chest X-ray, which was confirmed by ultrasound, highlighting a potential pitfall in AI-based radiographic interpretation.

## Contribution

Demonstrates a specific limitation of AI in detecting atypical pleural effusion patterns and emphasizes the need for ultrasound confirmation in such cases.

## Key findings

- AI correctly identified a bilateral pleural effusion but underestimated the right-sided subpulmonic effusion due to atypical imaging patterns.
- Thoracic ultrasound confirmed the presence of a right subpulmonic effusion that was not clearly visible on the chest radiograph.
- The case highlights the importance of using ultrasound to validate ambiguous or discordant AI findings in chest radiography.

## Abstract

Chest radiography remains an important exam for the initial assessment of suspected thoracic disease, including screening and orientation of patients in urgent care settings without on-site specialty coverage. In parallel, there has been a growing adoption of artificial intelligence to assist in the interpretation of chest radiographs; while not diagnostic on its own, it functions as an important clinical adjunct for prioritization and decision support. We report a case in which an artificial intelligence system correctly flagged a bilateral pleural effusion but assigned a lower probability to the right side because a subpulmonic distribution simulated hemidiaphragm elevation and apparently preserved the costophrenic angle on the posteroanterior projection, thereby lowering algorithmic confidence. Screening thoracic ultrasound confirmed a right subpulmonic effusion. The extent of pleural fluid was assessed visually (ocular estimation) and documented accordingly, explaining the apparent asymmetry despite bilateral pleural effusions.

This multimodal correlation prevented misclassification as isolated diaphragmatic elevation and informed subsequent management. This case illustrates a recognizable, discreet pitfall for artificial intelligence systems trained to detect more typical meniscus patterns and underlines the value of targeted ultrasound when radiographic signs are subtle, atypical, or discordant with clinical suspicion. A practical implication stems from the fact that when chest radiography suggests apparent elevation of the hemidiaphragm with a non-obliterated or partially maintained costophrenic angle, physicians should actively consider a subpulmonary distribution and confirm it with thoracic ultrasound, instead of relying on chest radiography alone, even when supported by AI.