Deep convolutional neural networks in the classification of dual-energy thoracic radiographic views for efficient workflow: analysis on over 6500 clinical radiographs

Abstract

DICOM header information is frequently used to classify medical image types; however, if a header is missing fields or contains incorrect data, the utility is limited. To expedite image classification, we trained convolutional neural networks (CNNs) in two classification tasks for thoracic radiographic views obtained from dual-energy studies: (a) distinguishing between frontal, lateral, soft tissue, and bone images and (b) distinguishing between posteroanterior (PA) or anteroposterior (AP) chest radiographs. CNNs with AlexNet architecture were trained from scratch. 1910 manually classified radiographs were used for training the network to accomplish task (a), then tested with an independent test set (3757 images). Frontal radiographs from the two datasets were combined to train a network to accomplish task (b); tested using an independent test set of 1000 radiographs. ROC analysis was performed for each trained CNN with area under the curve (AUC) as a performance metric. Classification between frontal images (AP/PA) and other image types yielded an AUC of 0.997 [95% confidence interval (CI): 0.996, 0.998]. Classification between PA and AP radiographs resulted in an AUC of 0.973 (95% CI: 0.961, 0.981). CNNs were able to rapidly classify thoracic radiographs with high accuracy, thus potentially contributing to effective and efficient workflow.

Toward understanding deep learning classification of anatomic sites: lessons from the development of a CBCT projection classifier

“…The number of training images per class for a given view angle was less than 500 (see Table 1), which is relatively small compared with reference datasets used to train models from scratch. 5,13 Therefore, transfer learning was utilized in this work. CNN hyperparameters, such as the optimizer, number of epochs, etc., were arbitrarily selected and kept fixed in this study.…”

“…While these features do not have an obvious interpretation, they could be due to subtle correlations in the training data, which may not be indicative of the class globally. 13,28,29 For example, it could be possible that image edges contain fingerprints of the hardware where the patient was scanned, which may correlate with the anatomic class since patients of the same treatment site are often grouped on the same treatment machine. Retrospective inspection of the projection headers ruled out the presence of imaging blades at the border of the projection images.…”

“…Others have shown similar findings noting that this may make translation to other data sets less effective. 13,28 Thus, to increase generalizability, the model development would, ideally, include multi-institutional data to reduce the impact of hardware or site-specific factors.…”

“…2.4). As others have done for medical image applications, 12,13 PNG images were rebinned to match the CNN architecture input size. 14…”

Machine Learning Augmented Interpretation of Chest X-rays: A Systematic Review

“…The quality of the dataset labeling method is likely to be a cornerstone of safe deep learning model development for systems intended for clinical use. Open source datasets may be vulnerable to adversarial perturbation, which can induce model failure or falsely high performance in image classification tasks [103]. Image perturbations are often difficult to detect.…”