Validation of a Breast Density Computer-aided Detection Software at Kenyatta National Hospital

Abstract

Background Mammographic breast density affects the sensitivity and specificity of mammography for the detection of breast masses. It is also an independent risk factor for breast cancer. Increased breast density is associated with an increased cost of supplemental screening. The assessment of mammographic breast density is highly subjective with significant Intra and interobserver variability well documented. There are several computer-assisted methods of breast density categorization with varying techniques used including area-based and volumetric categorization. Deep learning has emerged as a viable automated method of categorization with multiple studies showing acceptable agreement with expert radiologists. The majority of the available tools are developed using data generated abroad and are not validated in Kenya. It is crucial to validate any computer-aided detection tools locally to improve user confidence and acceptability. Aim To validate a locally-developed breast density classification deep learning model at Kenyatta National Hospital Methodology This was a cross-sectional study to determine the level of agreement between blinded radiology experts and a locally developed deep learning model on breast density classification using the American College of Radiology Breast Imaging Reporting and Data System. The mammograms evaluated in the study were a set of 400 anonymized digital mammograms stored in the KNH mammography section in 2021. The experts were three KNH radiologists with a combined experience of 22 years. Ethical approval and a waiver for informed consent were sought and approved given the retrospective nature of the data. Results: The findings from the study have shown the majority of mammograms from the KNH dataset were rated as heterogeneously dense (ACR C) up to 46.2% (n=400). There was fair to moderate agreement between expert radiologists' pairs with weighted kappa values ranging between 0.31 and 0.57 inclusive of the 95% confidence intervals. Fair to moderate agreement between individual radiologists and the AI model was demonstrated with weighted kappa values ranging from 0.34 - 0.56. Moderate agreement (weighted kappa 0.49 [0.42 - 0.56] was shown between the Experts' consensus and the AI model. The statistically significant results showed that the deep-learning model concordance values were within or above the range of experts’ concordance values. Conclusion This study has shown that the breast density validation deep learning model achieved professional-level performance