Deep Learning in Cardiothoracic Ratio Calculation and Cardiomegaly Detection

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Radiology

Background:

• Cardiothoracic ratio (CTR) is a key indicator for assessing cardiomegaly and pericardial effusion on chest radiography (CXR).
• Accurate CTR calculation is crucial for timely diagnosis and patient management.
• Current methods may be time-consuming or prone to inter-observer variability.

Purpose of the Study:

• To evaluate a deep learning algorithm's performance in calculating CTR from CXRs.
• To assess the algorithm's utility in identifying cardiomegaly and pericardial effusion.
• To compare AI-derived CTR measurements with human expert assessments.

Main Methods:

• A deep learning model was trained on 1020 randomly selected CXR images.
• The AI model segmented heart and lung anatomy to determine CTR.
• Measurements were compared against those obtained using standard DICOM viewer software and manual annotations.

Main Results:

• The AI model achieved high performance metrics, with Intersection over Union of 88.28% (training) and 83.06% (validation).
• F1-scores were 90.22% (training) and 90.67% (validation).
• AI-derived CTR, transverse thoracic diameter (TTD), and transverse cardiac diameter (TCD) showed statistically significant differences compared to human measurements (p < 0.001).

Conclusions:

• The deep learning algorithm demonstrates a significant correlation with human measurements for CTR calculation.
• The AI method shows potential as a screening tool or advisory support in clinical settings, particularly in time-sensitive environments like ICUs and ERs.
• Further validation in clinical conditions is recommended before widespread adoption.