Concurrent Computer-Aided Detection Improves Reading Time of Digital Breast Tomosynthesis and Maintains Interpretation Performance in a Multireader Multicase Study

Abstract

Concurrent use of CAD with DBT resulted in 29.2% faster reading time, while maintaining reader interpretation performance.

Digital Breast Tomosynthesis: Concepts and Clinical Practice

“…To address reader efficiency, new presentation modes, including automated "slabbing" of DBT sections to create thicker overlapping sections, have been developed (54). Machine learning-based detection algorithms have also been created to "flag" sections as well as to document the thickness of a detected lesion within the DBT stack to more rapidly guide DBT review (101). In addition, there are now U.S. Food and Drug Administration-approved DBT machine-learning computer-aided detection programs that use lesion detection data to drive the generation of synthetic images.…”

“…In addition, there are now U.S. Food and Drug Administration-approved DBT machine-learning computer-aided detection programs that use lesion detection data to drive the generation of synthetic images. These "computer-aided detection-enhanced" images increase the conspicuity of lesions on SM images, and in early reader studies have been shown to decrease reading time by 29.2% while maintaining performance accuracy (101). The use of machine learning coupled with computer-aided detection is also being investigated to quantitatively classify breast density categories, predict benign versus malignant masses, and improve the diagnostic accuracy and efficiency of radiologists (102)(103)(104).…”

Interpretation time for screening mammography as a function of the number of computer-aided detection marks

“…A survey of 400 mammography sites in 2007 was followed up by surveys of the same sites in 2011 and 2016, and each time the percentage of sites using CAD for screening mammography was just over 90%, ranging from 90.2% in 2011 to 92.3% in 2016. 9 CAD systems are also now being developed for use with digital tomosynthesis, 10 so even though conventional mammography is becoming less frequent as a stand-alone study, information on how humans interact with CAD systems remains relevant.…”

CANDI: an R package and Shiny app for annotating radiographs and evaluating computer-aided diagnosis

“…Algorithms designed for Computer-Aided Diagnosis (CAD) are frequently only evaluated in isolation, and studies evaluating human performance with and without CAD have had inconsistent results. Retrospective studies on engineered feature (not DL) CAD in clinical practice have found accuracy benefit (Kasai et al , 2008), no accuracy benefit (Benedikt et al , 2017), or a negative effect (Gilbert et al , 2008). CAD enhancement of human interpretation has been studied in disparate experimental designs.…”

“…Commercially available CAD tools have been tested in fully randomized studies (Gilbert et al , 2008) and observational studies (Fenton et al , 2007). Experimental algorithms have been tested in only one mode (see RCT Case Study below) (Kasai et al , 2008), or over multiple sessions (double-crossover design) where one day a radiologist interprets images with CAD and several months later she interprets images without CAD (or vice versa, by randomization) (Benedikt et al , 2017). RCTs are graded as stronger evidence than pseudorandomized or observational studies, but RCTs have only been done with commercially available CAD systems.…”