Duty of Candour in Interval Cancer Reviews: Does AI add Transparency?

William Teh, Mia Morgan

Objective:
To evaluate an AI system for detection of interval cancers (IC) in the NHS breast screening programme (NHSBSP), and its potential to be used by radiologists for support during the review process of IC cases.

Methods:
Screening exams of the 232 interval cancers of the NHSBSP program were consecutively collected. Each IC case was categorized following NHS guidelines into three categories: „satisfactory IC“ – when the exam at screening shown no reason to recall; „satisfactory with learning points“ – when the exam at screening showed the lesion only with hindsight, it was difficult to perceive and may provide learning points; and „unsatisfactory“ – when the exam at screening had a radiological appearance obviously malignant and should have been recalled. Each screening exams of IC was analyzed with a CAD-AI cancer detection system, Transpara 1.6.0 (ScreenPoint Medical), yielding a suspicion score 1-10 representing the likelihood of cancer in the exam. Breast density assessments were also recorded following BIRADS guidelines.

Results:
167/232 (72%), 54/232 (23%), and 11/232 (4.7%) of IC cases were categorized during the review process as category 1, 2 and 3 respectively. There was a strong association between the category of IC and breast density (58% of cases with density C or D were IC type 1, while 64% of cases with density A or B were IC type 3). All IC cases with category 3, had an AI score of 9 or 10. For IC cases with category 2, 77% had an AI score of 9 or 10. The majority of cases with AI scores 1 to 5 were IC cases of category 1. Nevertheless, a significant number of occult / true interval cancers (type 1) got assigned a high AI score of 10.

Conclusion:
AI could be used for support during the review process of IC cases to provide a more homogeneous categorization of IC types, but further studies are needed as well as technological improvements such as the analysis of prior exams.