External Validation of a Deep Learning Based Model for Pulmonary Embolism Detection on Chest CT Scans
External Validation of a Deep Learning Based Model for Pulmonary Embolism Detection on Chest CT Scans
The scope of this study is the external validation of an explainable deep learning-based classifier for the diagnosis and detection of pulmonary embolism in computed tomography pulmonary angiography (CTPA) and contrast enhanced CT scans.
Pulmonary embolism (PE) is a potentially fatal disease if not promptly diagnosed and treated. Chest CTPA remains the gold standard for diagnosis nowadays, but PE can also be incidentally found on enhanced CT scans. Most CTPA exams are performed in clinics in case of suspicion of PE in urgent conditions, whereas a minority is performed for conditions of suspicious or validated chronic pulmonary thromboembolism, a disease frequently overlooked on CT scans but affected by high morbidity and poor prognosis if left untreated. Thus methods to expedite and automatize the recognition of emboli within pulmonary vessels have the potential of becoming an important support in clinical practice, enabling the better triage of urgent cases of PE and an increased sensitivity in the identification of patients with chronic pulmonary thromboembolism. Based on these clinical needs, a deep learning-based model for the detection of pulmonary embolism has been developed on CTPA scans. The model was based on 2D ResNext50 architecture and was trained and validated using a multicentric open source dataset composed of 7169 patients. From these retrospective data, 85,000 slices positive for PE and 123,428 negative for PE were extracted for training. For internal validation, 9,922 slices were used for each class. The model was initially externally validated at the patient-level using a dataset of 156 adult patients from 3 different public sources, with all emboli segmented by at least one experienced radiologist. To gain insight into the model predictions, activation maps were extracted using the Grad-CAM method. Comparing these maps with the ground truth (GT) segmentations, it was determined if the activated regions corresponded to regions of PE by computing the percentage of GT PE that was activated and the percentage of activated regions corresponding to GT PE. The PE classification model reached an area under the curve (AUC) of 0.86 [0.800-0.919], a sensitivity of 82.68 % [75.16 - 88.27] and a specificity of 79.31 % [61.61 - 90.15] on the external validation set. However, these results have been obtained in an unbalanced external validation cohort (127 PE positive against 29 PE negative patients), thus it is very important to assess the model performances also in a more balanced patients cohort, representing the real clinical incidence of PE (between 12 and 22%). For this reason the scope of the present study is to collect an external validation cohort representative of the real clinical reality, including both CTPA and enhanced CT scans, with a more balanced percentage of positive and negative PE cases. Moreover, the performances of the model will be compared between enhanced CT and CTPA scans.
Inclusion Criteria:
Exclusion Criteria: