T. J. Weikert, D. J. Winkel, J. Bremerich, B. Stieltjes, A. W. Sauter, G. Sommer; Basel/CH
Purpose: To validate the performance of deep convolutional neural networks optimised for the detection of pulmonary embolism (PE) on CT pulmonary angiograms (CTPAs).
Methods and Materials: We downloaded all CTPAs performed in 2017 along with the corresponding reports (n = 1,499) from our RIS/PACS archive using an in-house-developed search engine. The reports were manually reviewed by a radiologist. CTPAs with other clinical questions than PE or poor diagnostic quality were excluded. The remaining exams were then classified into positive (n = 232) and negative (n = 1,204) for PE. All emboli in positive exams were labeled by a radiologist using bounding boxes. The data served as ground truth for the validation of prototype algorithms (Aidoc, Tel Aviv, Israel) that had previously been trained on 28.000 independent CTPAs from other centres.
Results: Four trained prototype algorithms were tested on our CTPA dataset. The best performing algorithm was a fully convolutional neural network with a backbone based on the Resnet architecture. It achieved a sensitivity of 93% and a specificity of 95%. This corresponds to a positive predictive value of 77%.
Conclusion: The best-performing AIalgorithm we validated is capable of detecting pulmonary embolism in CTPAs with a high sensitivity and specificity. In a clinical setting, this can complement conventional workflows with a worklist prioritisation and has the potential to improve the quality of healthcare by accelerating the diagnostic process and communication. We plan to further test the algorithm and finally implement it in the clinical routine to perform prospective evaluations.