Pilot study: Application of artificial intelligence for detecting left atrial enlargement on canine thoracic radiographs

Abstract Although deep learning has been explored extensively for computer‐aided medical imaging diagnosis in human medicine, very little has been done in veterinary medicine. The goal of this retrospective, pilot project was to apply the deep learning artificial intelligence technique using thoracic radiographs for detection of canine left atrial enlargement and compare results with those of veterinary radiologist interpretations. Seven hundred ninety‐two right lateral radiographs from canine patients with thoracic radiographs and contemporaneous echocardiograms were used to train, validate, and test a convolutional neural network algorithm. The accuracy, sensitivity, and specificity for determination of left atrial enlargement were then compared with those of board‐certified veterinary radiologists as recorded on radiology reports. The accuracy, sensitivity, and specificity were 82.71%, 68.42%, and 87.09%, respectively, using an accuracy driven variant of the convolutional neural network algorithm and 79.01%, 73.68%, and 80.64%, respectively, using a sensitivity driven variant. By comparison, accuracy, sensitivity, and specificity achieved by board‐certified veterinary radiologists was 82.71%, 68.42%, and 87.09%, respectively. Although overall accuracy of the accuracy driven convolutional neural network algorithm and veterinary radiologists was identical, concordance between the two approaches was 85.19%. This study documents proof‐of‐concept for application of deep learning techniques for computer‐aided diagnosis in veterinary medicine.

left atrial enlargement, an early feature of some dogs with myxomatous mitral valve disease, is therefore essential for appropriate initial medical management and to assess risk of heart failure and prognosis. [3][4][5][6][7] A presumptive diagnosis of myxomatous mitral valve disease is often reached based on signalment, a left apical systolic murmur, and characteristic thoracic radiographic features. It is confirmed with an echocardiographic examination. Echocardiography is relatively expensive, requires specialized training to perform accurately, and is of limited access in general practice. Currently, the clinically applicable methods to identify left atrial enlargement from thoracic radiographs include detection of characteristic cardiac margin changes, carinal elevation, subjective mainstem bronchial widening, bifurcation angle measurements, and vertebral heart score estimations. [8][9][10] None of these, however, are considered consistently accurate, and radiographic interpretation of left atrial enlargement has been shown to be inconsistent, particularly when performed by those without advanced training. 11 Thoracic radiographic examination is readily accessible and easily performed, but its value as a test for assessing left atrial enlargement varies depending on the interpretive skill of the clinician evaluating the study and the accuracy of thoracic radiographs for assessing left atrial enlargement. The use of automated diagnosis from radiographic images using deep learning, an artificial intelligence technique that may match or exceed human expert performance in recognition of highly heterogeneous diagnostic images, has recently gained traction. 12,13 Indeed, the number of peer-reviewed articles published annually relating to convolutional neural networks or deep learning have increased exponentially over the past 5 years. 14,41 Image analysis using deep learning techniques has been used in human medicine with success to detect and stage diabetic retinopathy and to accurately differentiate radiographs of patients with tuberculosis from normal controls given a sufficiently large training dataset. 12,15 The specific deep learning approach best suited for analyzing diagnostic images is the application of convolutional neural network. 16 Convolutional neural network is essentially a complex computer algorithm that is commonly used for image analysis. The algorithm contains multiple processing layers and many parameters; however, it is not explicitly programmed or pre-defined, thus referred as "black box." The parameters within the algorithm were tuned to achieve a good fit between the input (images) and classification labels (in our study, left atrial enlargement). Input data go through multiple processing layers of abstraction to match the label. When the algorithm is exposed to large amount of data, mathematically, it has higher chance to give more accurate results.
Based on our review of the literature, a convolutional neural network model has not been explored in veterinary medicine as a means of assisting imaging diagnosis but has the potential to be an affordable, rapid, and reliable tool for veterinary medical diagnosis. The purpose of this investigation was to create an automated imaging diagnostic tool using deep learning techniques and test its accuracy to that of veterinary radiologists.

Dataset
Medical records from the University of California, Davis, Veterinary Medical Teaching Hospital (VMTH) from 2010 to 2017 were screened for canine patients who had a thoracic radiographic examination and a contemporaneous echocardiographic examination performed within 72 hours of the radiographic study. All radiographs were acquired using the same radiographic units with DICOM output Digital radiog- Images were designated as being echocardiographically "positive" or "negative" for left atrial enlargement based on conclusions in corresponding echocardiographic reports. Similarly, images were designated radiographically positive or negative based on the corresponding radiology reports.

Methods
This was a retrospective pilot study. Right lateral thoracic radiographic DICOM images were downloaded as .jpeg files directly from the hospital PACS server with no initial alteration of native matrix size. dataset images with the prediction results averaged from these 10 models. The prediction results were presented as a probability, a number between 0 and 1, indicating the likelihood that a given image was positive for left atrial enlargement. When the P was >.5, the result was interpreted as a "positive prediction." The results of the testing dataset were used to assess the diagnostic accuracy, sensitivity, and specificity.  Table 1).
In addition, we cropped the heart regions from all original images using the following criteria: visible cranial margin of the heart, visible caudal margin of the heart, ventral margin of the spine, and dorsal margin of the sternum. We performed our analysis using the same techniques and models. The results are very similar to the one using the entire image size. The detailed results revealed that there is almost no difference between two approaches using uncropped and cropped data.  Table 2). The overall accuracy of the accuracy driven convolutional neural network model was 82.71% with a sensitivity of 68.42% and specificity of 87.09% (Table 6).

Seven hundred ninety
In the sensitivity driven convolutional neural network model, of the 81 images in the testing dataset, 14 positive images were predicted positive, 5 positive images were predicted negative, 12 negative images were predicted positive, and 50 negative images were predicted negative using the sensitivity driven convolutional neural network model ( Table 3). The overall accuracy of the sensitivity driven convolutional neural network model was 79.01% with a sensitivity of 73.68% and specificity of 80.64% (Table 6).
For board-certified veterinary radiologists, of the 792 images in the entire data set, 208 positive images were predicted positive, 73 positive images were predicted negative, 64 negative images were predicted positive, and 447 negative images were predicted negative based on radiologists radiographic reports (Table 4).

Predict positive Predict negative Total
True Positive 13 6 19 True Negative  8  54  62   Total  21  60  81 13 positive images were predicted positive, 6 positive images were predicted negative, 8 negative images were predicted positive, and 54 negative images were predicted negative based on radiologists radiographic reports ( Table 5). The accuracy of board-certified radiologists interpretation of the testing dataset was 82.71% with a sensitivity of 68.42% and specificity of 87.09%. (Table 6). Table 6 shows the comparison of performance of the accuracy-

DISCUSSION
In this preliminary investigation, the convolutional neural network model was trained and validated using single right lateral radiographic images down-sized to a 64 × 64 matrix size to keep the computational analysis to a reasonable scale given the available computing resources.
Despite the fact that these data were compared to veterinary radiologist interpreted examinations that included three-view, high resolution images and a pertinent patient history, the convolutional neural network model was able to consistently achieve similar accuracy and sometimes higher sensitivity in detection of left atrial enlargement.
These findings are consistent with a recent meta-analysis investigat-  ing study as "high-risk" for follow-up review by a specialist, serving as a second "over-read" following a primary clinician or specialist interpretation or simply serving as a fully automated screening or diagnostic test. Although the latter alternative is not currently applicable, rapid advances in the field of artificial intelligence and computeraided diagnosis may make automated diagnosis feasible in the near future. [37][38][39][40] A number of clinically applicable approaches have been used to assess left atrial enlargement from thoracic radiographs including subjective evaluation of cardiac contours, tracheal bifurcation angle measurements, and vertebral left atrial size. 7,8 Yet none of these methods has been proven to be consistently accurate. This is, in large part, due to inherent limitations of the thoracic radiographic examination for assessment of left atrial enlargement. Confounding variations in the appearance of the cardiac silhouette are caused by breed variability, patient positioning differences, cardiac and respiratory phase, and cardiac and noncardiac co-morbidities, among other parameters. Inconsistent assessment can also be the result of inter-reader or intra-reader interpretation variability. The use of echocardiographic findings as a standard for left atrial size has its own inherent limitation as the inaccuracy from echocardiograph will be transferred into the convolutional neural network model. One big potential problem of the convolutional neural network is the overfitting, 19 in particular when the dataset is small. The dropout layer technique is applied to reduce it. The convolutional neural network model described in this study however provides an objective evaluation tool that can consistently improve itself by continually expanding the training dataset of echocardiograph-validated images. The application of deep learning in detecting left atrial enlargement using this approach could potentially improve detection accuracy and possibly reduce "time to diagnosis" through automation.
As with any study using retrospectively derived data, there are a number of limitations to this investigation. Matrix down-sizing, necessary to distill imaging data down to a manageable computational size, necessarily reduces quality of the data as does limiting analysis to a single right lateral view. We also recognize that changes in left atrial size could occur within 72 hours separating radiographic and echocardiographic examinations. Despite these limitations, as the amount of data used to train the model increases; including number of patients, the number of image views, and image matrix size; the convolutional neural network model might be expected to achieve higher accuracy, sensitivity, and specificity in the future.

CONCLUSION
Results from this preliminary investigation suggest that the convolutional neural network model is applicable to canine thoracic image analysis and computer-aided diagnosis. The convolutional neural network model was shown to achieve accuracy and sensitivity similar to veterinary radiologists in detecting left atrial enlargement and has the potential to be modified to address other clinically pertinent diagnostic questions.