Room: Exhibit Hall | Forum 1
Purpose: Automated quality assurance of chest x-rays can improve the quality and consistency of images, decrease repeat rates and reduce patient dose. The goal of this work is to determine if convolutional network based lung segmentation models with an additional classification algorithm can accurately classify clipped images in a clinical setting.
Methods: A U-Net convolutional network was trained on 274 images from the JSRT chest x-ray data set (â€œBase Modelâ€?). A second model was trained based on this â€œBase Modelâ€? with 18 images from the NIH Chest dataset that were contoured by hand (â€œUpdated Modelâ€?). Both models were applied to 1000 randomly selected images from NIH dataset, which were manually tagged as â€œclippedâ€? and â€œnot-clippedâ€?. For each test image, each model generated a segmentation map of the lung region, and the image was classified as â€œclippedâ€? and â€œnot-clippedâ€? based on this segmentation map. Model predictions were compared with manual tagging and the results were analyzed.
Results: Both â€œBase modelâ€? and â€œUpdated Modelâ€? performed reasonably well on the majority of the images with an accuracy rate of 0.91 and 0.94, respectively. The â€œUpdated Modelâ€? has a slightly decreased segmentation accuracy (Dice coefficients reduced from 0.985 to 0.969). Compared with the â€œBase Modelâ€?, the â€œUpdated Modelâ€? had much higher accuracy and sensitivity but lower specificity. Overall, 6.1% of the images were identified as â€œclippedâ€? in the 1000 test images.
Conclusion: Machine learning methods were developed to automatically detect clipped lung images in clinical chest x-rays with high accuracy. A promising improvement was implemented to adapt the model to the clinical challenging cases with a limited number of new training data.
Chest Radiography, Quality Control, Computer Vision