Purpose To determine whether saliency maps in radiology artificial intelligence (AI) are vulnerable to subtle perturbations of the input, which could lead to misleading interpretations, using prediction-saliency correlation (PSC) for evaluating the sensitivity and robustness of saliency methods. Materials and Methods In this retrospective study, locally trained deep learning models and a research prototype provided by a commercial vendor were systematically evaluated on 191 229 chest radiographs from the CheXpert dataset and 7022 MR images from a human brain tumor classification dataset. Two radiologists performed a reader study on 270 chest radiograph pairs. A model-agnostic approach for computing the PSC coefficient was used to evaluate the sensitivity and robustness of seven commonly used saliency methods. Results The saliency methods had low sensitivity (maximum PSC, 0.25; 95% CI: 0.12, 0.38) and weak robustness (maximum PSC, 0.12; 95% CI: 0.0, 0.25) on the CheXpert dataset, as demonstrated by leveraging locally trained model parameters. Further evaluation showed that the saliency maps generated from a commercial prototype could be irrelevant to the model output, without knowledge of the model specifics (area under the receiver operating characteristic curve decreased by 8.6% without affecting the saliency map). The human observer studies confirmed that it is difficult for experts to identify the perturbed images; the experts had less than 44.8% correctness. Conclusion Popular saliency methods scored low PSC values on the two datasets of perturbed chest radiographs, indicating weak sensitivity and robustness. The proposed PSC metric provides a valuable quantification tool for validating the trustworthiness of medical AI explainability. Keywords: Saliency Maps, AI Trustworthiness, Dynamic Consistency, Sensitivity, Robustness Supplemental material is available for this article. © RSNA, 2023 See also the commentary by Yanagawa and Sato in this issue.

Chest radiographs (CXR) are the most performed imaging tests and rank high among the radiographic exams with suboptimal quality and high rejection rates. Suboptimal CXRs can cause delays in patient care and pitfalls in radiographic interpretation, given their ubiquitous use in the diagnosis and management of acute and chronic ailments. Suboptimal CXRs can also compound and lead to high inter-radiologist variations in CXR interpretation. While advances in radiography with transitions to computerized and digital radiography have reduced the prevalence of suboptimal exams, the problem persists. Advances in machine learning and artificial intelligence (AI), particularly in the radiographic acquisition, triage, and interpretation of CXRs, could offer a plausible solution for suboptimal CXRs. We review the literature on suboptimal CXRs and the potential use of AI to help reduce the prevalence of suboptimal CXRs.

Background: Missed findings in chest X-ray interpretation are common and can have serious consequences. Methods: Our study included 2407 chest radiographs (CXRs) acquired at three Indian and five US sites. To identify CXRs reported as normal, we used a proprietary radiology report search engine based on natural language processing (mPower, Nuance). Two thoracic radiologists reviewed all CXRs and recorded the presence and clinical significance of abnormal findings on a 5-point scale (1—not important; 5—critical importance). All CXRs were processed with the AI model (Qure.ai) and outputs were recorded for the presence of findings. Data were analyzed to obtain area under the ROC curve (AUC). Results: Of 410 CXRs (410/2407, 18.9%) with unreported/missed findings, 312 (312/410, 76.1%) findings were clinically important: pulmonary nodules (n = 157), consolidation (60), linear opacities (37), mediastinal widening (21), hilar enlargement (17), pleural effusions (11), rib fractures (6) and pneumothoraces (3). AI detected 69 missed findings (69/131, 53%) with an AUC of up to 0.935. The AI model was generalizable across different sites, geographic locations, patient genders and age groups. Conclusion: A substantial number of important CXR findings are missed; the AI model can help to identify and reduce the frequency of important missed findings in a generalizable manner.

Purpose: Motion-impaired CT images can result in limited or suboptimal diagnostic interpretation (with missed or miscalled lesions) and patient recall. We trained and tested an artificial intelligence (AI) model for identifying substantial motion artifacts on CT pulmonary angiography (CTPA) that have a negative impact on diagnostic interpretation. Methods: With IRB approval and HIPAA compliance, we queried our multicenter radiology report database (mPower, Nuance) for CTPA reports between July 2015 and March 2022 for the following terms: “motion artifacts”, “respiratory motion”, “technically inadequate”, and “suboptimal” or “limited exam”. All CTPA reports were from two quaternary (Site A, n = 335; B, n = 259) and a community (C, n = 199) healthcare sites. A thoracic radiologist reviewed CT images of all positive hits for motion artifacts (present or absent) and their severity (no diagnostic effect or major diagnostic impairment). Coronal multiplanar images from 793 CTPA exams were de-identified and exported offline into an AI model building prototype (Cognex Vision Pro, Cognex Corporation) to train an AI model to perform two-class classification (“motion” or “no motion”) with data from the three sites (70% training dataset, n = 554; 30% validation dataset, n = 239). Separately, data from Site A and Site C were used for training and validating; testing was performed on the Site B CTPA exams. A five-fold repeated cross-validation was performed to evaluate the model performance with accuracy and receiver operating characteristics analysis (ROC). Results: Among the CTPA images from 793 patients (mean age 63 ± 17 years; 391 males, 402 females), 372 had no motion artifacts, and 421 had substantial motion artifacts. The statistics for the average performance of the AI model after five-fold repeated cross-validation for the two-class classification included 94% sensitivity, 91% specificity, 93% accuracy, and 0.93 area under the ROC curve (AUC: 95% CI 0.89–0.97). Conclusion: The AI model used in this study can successfully identify CTPA exams with diagnostic interpretation limiting motion artifacts in multicenter training and test datasets. Clinical relevance: The AI model used in the study can help alert technologists about the presence of substantial motion artifacts on CTPA, where a repeat image acquisition can help salvage diagnostic information.

Pulmonary embolism is a common cause of cardiopulmonary mortality and morbidity worldwide. Survivors of acute pulmonary embolism may experience dyspnea, report reduced exercise capacity, or develop...