A computer vision-based algorithm to predict false positive errors in radiology trainees when interpreting digital breast tomosynthesis cases. (1st December 2016)

• Record Type: Journal Article
• Authors: Wang, Mengyu; Wang, Meng; Grimm, Lars J.; Mazurowski, Maciej A.
• Is Part Of: Expert systems with applications. Volume 64(2016)
• Page Start: 490
• Page End: 499
• Publication Date: 2016-12-01
• Subjects: Radiology education -- False positive error prediction -- Training plan optimization -- Digital breast tomosynthesis -- Image processing -- Clustering; Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2016.08.023
• Languages: English
• ISSNs: 0957-4174

Abstract

Highlights: We developed an algorithm to identify trainee's false positive errors on DBT. The algorithm is personalized for each individual trainee. Our model can improve trainees' training by focusing on false positive error cases. Abstract: Objectives: Digital breast tomosynthesis (DBT) is a new imaging modality that improves invasive cancer detection rates compared to mammography. In this work, we aim to advance adaptive computer-based education in DBT by computer algorithm. Methods: First, a set of potentially difficult locations are identified based on locations marked by other trainees using a regional clustering algorithm. Second, the candidate location is segmented to identify potential abnormal objects. Third, 18 features are extracted from the location from the segmented image. Finally, a classifier uses the 18 features to predict whether the candidate location would result in a false positive error for the trainee. The classifier is personalized for each trainee by using data from the trainee's prior DBT interpretations. Results: Our algorithm successfully identified locations more likely associated with false positive errors as compared to randomly identified locations. The prevalence of errors among the difficult locations was 20.7% when 1 location per trainee was predicted and 17.2% when 10 locations were predicted. In comparison, the prevalence of errors for random locations generated within a breast region with 1 and 10 identified locations was 0% and 4.8%, respectively. Conclusions: We developed an algorithm to successfully identify locations on DBT where trainees are more likely to commit false positive errors. Advances in knowledge: Our user model can be used to select the most challenging cases for each trainee from the perspective of committing false positive errors. Our model improved the status quo of case presentation with random selection to trainee in breast tomosynthesis.