AI-aided detection of malignant lesions in mammography screening – evaluation of a program in clinical practice

Abstract

Objectives: Evaluation of the degree of concordance between an artificial intelligence (AI) program and radiologists in assessing malignant lesions in screening mammograms. Methods: The study population consisted of all consecutive cases of screening-detected histopathologically confirmed breast cancer in females who had undergone mammography at the NU Hospital Group (Region Västra Götaland, Sweden) in 2018 to 2019. Data were retrospectively collected from the AI program (lesion risk score in percent and overall malignancy risk score ranging from 1 to 10) and from medical records (independent assessments by two radiologists). Ethical approval was obtained. Results: Altogether, 120 females with screening-detected histopathologically confirmed breast cancer were included in this study. The AI program assigned the highest overall malignancy risk score 10 to 86% of the mammograms. Five cases (4%) were assigned an overall malignancy risk score ≤5. Lack of consensus between the two radiologists involved in the initial assessment was associated with lower overall malignancy risk scores (p = 0,002). Conclusion: The AI program detected a majority of the cancerous lesions in the mammograms. The investigated version of the program has, however, limited use as an aid for radiologists, due to the pre-calibrated risk distribution and its tendency to miss the same lesions as the radiologists. A potential future use for the program, aimed at reducing radiologists’ workload, might be to preselect and exclude low-risk mammograms. Although, depending on cut-off score, a small percentage of the malignant lesions can be missed using this procedure, which thus requires a thorough risk–benefit analysis. Advances in knowledge: This study conducts an independent evaluation of an AI program’s detection capacity under screening-like conditions which has not previously been done for this program.

Barriers and facilitators of artificial intelligence conception and implementation for breast imaging diagnosis in clinical practice: a scoping review

“…Related to the interpretability problem, the adoption of an AI system depends on the level of confidence that the radiologists have in the AI system and its findings, and also on the understanding they have about the tool ( n = 19) [ 24 , 33 , 34 ]. Avoiding dependence on AI systems to preserve users’ ability to think critically and make good patient decisions is also linked to radiologists’ trust [ 35 , 36 ].…”

“…Clinician trust (B2.2) Related to the interpretability problem, the adoption of an AI system depends on the level of confidence that the radiologists have in the AI system and its findings, and also on the understanding they have about the tool (n = 19) [24,33,34]. Avoiding dependence on AI systems to preserve users' ability to think critically and make good patient decisions is also linked to radiologists' trust [35,36]. Patient trust (B2.3) Patient confidence (n = 11) in the AI systems was also reported as important in the same way as for radiologists, and is also linked to communication between human and machine when the diagnosis is disclosed [3,37].…”

Artificial Intelligence in Breast Imaging: Opportunities, Challenges, and Legal–Ethical Considerations

“…Data quality relies on proper labeling, cleaning, and objective processing, with reliable labeling and automation enhancing accuracy. 55 , 65 , 66 Anonymization and security measures are crucial for addressing privacy concerns in data sharing. 56 , 63 , 67 - 73 …”

Leveraging law and ethics to promote safe and reliable AI/ML in healthcare

“…Omission errors refer to when a physician fails to detect an error in the CDS, such as an incorrect treatment recommendation, prescription dose, wrong drug not indicated by clinical guidelines, or other failure to detect an anomaly. As one example, AI/ML tools that use mammography to detect breast cancer could miss lesions, or misclassify malignant lesions as benign (29). Lyell et al classifies commission errors, on the other hand, as when physicians comply with incorrect recommendations or accept false positive alerts (15).…”