Abstract B078: Artificial-intelligence-driven breast density assessment in the transition from full-field digital mammograms to digital breast tomosynthesis

Abstract

Listen

Translate article

Abstract Introduction: To enhance reproducibility and robustness in mammographic density assessment, various artificial intelligence (AI) models have been proposed to automatically classify mammographic images into BI-RADS density categories. Despite their promising performances, so far density AI models have been assessed primarily in traditional full-field digital mammography (FFDM) images. Our study aims to assess the potential of AI in breast density assessment in FFDM versus the newer synthetic mammography (SM) images acquired with digital breast tomosynthesis. Methods: We retrospectively analyzed negative (BI-RADS 1 or 2) routine mammographic screening exams (Selenia or Selenia Dimensions; Hologic) acquired at sites within the Barnes-Jewish/Christian (BJC) Healthcare network in St. Louis, MO from 2015 to 2018. BI-RADS breast density assessments of radiologists were obtained from BJC’s mammography reporting software (Magview 7.1). For each mammographic imaging modality, a balanced dataset of 4,000 women was selected so there were equal numbers of women in each of the four BI-RADS density categories, and each woman had at least one mediolateral oblique (MLO) and one craniocaudal (CC) view per breast in that mammographic imaging modality. Previously validated pre-processing steps were applied to all FFDM and SM images to standardize image orientation and intensity. Images were then split into training, validation, and test sets at ratios of 80%, 10%, and 10%, respectively, while maintaining the distribution of breast density categories and ensuring that all images of the same woman appear only in one set. Our AI model was based on the widely used ResNet50 architecture and was designed to accept as an input a mammographic image and predict the BI-RADS breast density category that the image belongs to. Our AI model was optimized, trained, and evaluated separately for each mammographic imaging modality. We report on the AI model’s predictive accuracy on the test set for each mammographic imaging modality, for both views as well as separately for CC and MLO; accuracy differences in FFDM versus SM were assessed via bootstrapping. Results: A batch size of 32, learning rate of e-6, and Adam optimizer were chosen as the optimal hyperparameters for our AI model. Using the same hyperparameters, the AI model demonstrated substantially higher accuracy on the test set for FFDM than for SM (FFDM: accuracy = 71% ± 4.5% versus SM: accuracy = 66% ± 4.2%; p-value<0.001 for comparison). Similar conclusion held when CC and MLO views were evaluated separately (accuracy = 72% ± 4.6% versus 66% ± 4.3% for CC; accuracy = 69% ± 4.5% versus 62% ± 4.3% for MLO; p-value<0.001 for both comparisons). Conclusions: AI performance in BI-RADS breast density assessment was significantly higher on FFDM versus SM, even under the same AI model design, dataset size and training process. Our preliminary findings suggest that further AI optimizations and adaptations may be needed as we translate AI models from FFDM to the newer SM format acquired with digital breast tomosynthesis. Citation Format: Krisha Anant, Juanita Hernandez Lopez, Debbie Bennett, Aimilia Gastounioti. Artificial-intelligence-driven breast density assessment in the transition from full-field digital mammograms to digital breast tomosynthesis [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Advances in Breast Cancer Research; 2023 Oct 19-22; San Diego, California. Philadelphia (PA): AACR; Cancer Res 2024;84(3 Suppl_1):Abstract nr B078.