And when the CADx software is improved with an artificial intelligence algorithm, radiologist performance improves as well, noted presenter Dr. Wendie Berg, PhD, of the University of Pittsburgh School of Medicine, and colleagues.
Berg's group evaluated the effect of CADx in terms of improved sensitivity and specificity on radiologists' breast ultrasound interpretation. The study included 319 lesions identified on ultrasound screening, 27.6% of which were cancerous. Nine radiologists interpreted the images, reviewing them in random order and assigning them a BI-RADS category both with and without the use of CADx software.
The software was used in three modes: native CADx, high sensitivity, and high specificity. The researchers used the area under the curve (AUC) to assess the performance of the different modes.
CADx alone had an AUC of 0.82. The team also found the following:
- In native CADx mode, the average AUC was 0.82 without the use of CADx and 0.82 with CADx. In this mode, five readers had slight increase in AUC and four had a slight decrease, neither of which were statistically significant.
- In high-sensitivity mode, all readers' AUCs increased after the CADx cues: Average AUC was 0.83 before the use of CADx and 0.88 after (p = 0.0002).
- In high-specificity mode, all readers' AUCs increased after CADx cues: Average AUC was 0.82 before the use of CADx and 0.89 after (p < 0.0001).
The study results suggest that CADx can help improve radiologists' breast ultrasound reading performance, according to Berg and colleagues.
A radiologist's performance interpreting breast ultrasound can be improved by CADx if it performs at an AUC approximately 0.08 higher than the radiologist alone, the group concluded.
