Abstracts

RETURN TO ABSTRACT LISTING


E1843. Automated Calculation of Cardiothoracic Ratio From Chest Radiographs for Detection of Cardiomegaly
Authors
  1. Amit Kharat; DeepTek Imagng Private Limited
  2. Tanveer Gupte; DeepTek Imagng Private Limited
  3. Priyam Choudhury; DeepTek Imagng Private Limited
  4. Ashrika Gaikwad; DeepTek Imagng Private Limited
  5. Viraj Kulkarni; DeepTek Imagng Private Limited
  6. Aniruddha Pant; DeepTek Imagng Private Limited
Objective:
Cardiomegaly is a symptom of the underlying diseased condition of the heart, and assessing the heart size continues to be one of the important aspects on chest radiographs (1). However, its detection on chest X-ray (CXR) is highly subjective and dependent on radiologists' observational skills and experience. Cardiothoracic ratio (CTR) is used as the standard for measuring heart size (2). Therefore, an automated deep convolution neural networks (dCNN) model was designed, to compute the CTR, for an objective determination of the presence/absence of cardiomegaly to reduce the subjectivity.

Materials and Methods:
Densenet-121 architecture was used to differentiate between Antero-posterior (AP) / Postero-anterior (PA) CXR, as AP view may give a false impression of cardiomegaly due to larger heart shadow (3). 44,000 AP and 56,000 PA, pre-annotated CXR from the NIH (National Institutes of Health) database were used to design this first model. Mask R-CNN architecture was used to design a second model, to determine the CTR. 1420 CXR (710 AP and 710 PA) was annotated by radiologists using bounding boxes for heart and lungs.1000 test, 220 validation, and 200 test sets were used, to train the dCNN to segment the heart and lungs. Then, the network predicted a region of interest with bounding coordinates for both heart and chest. These coordinates were used for the determination of the width of the chest and heart and used to calculate the CTR to determine the presence of Cardiomegaly. CTR measurement above 0.50 was considered as cardiomegaly (4).

Results:
The segmentation models performed with the Dice score of 0.91 for the heart and 0.95 for the lungs. For the final classification into the presence or absence of cardiomegaly, accuracy and precision of 0.86, sensitivity of 0.83, and specificity of 0.88 were obtained for the PA view.

Conclusion:
Automated computing of CTR from CXRs in PA view makes the assessment of cardiomegaly more objective, accurate and faster, improving report turnaround times. The successful distinction between AP and PA CXR reduces any false cardiomegaly reporting from AP view, due to larger heart shadow (4).