Radiomics-based analysis of pre-symptomatic screening mammograms associated with future breast cancer occurrence in a Japanese screening population.
Authors
Affiliations (5)
Affiliations (5)
- Department of Radiological Technology, Tsukuba Medical Center Hospital, Tsukuba, 305-8558, Japan. [email protected].
- Graduate School of Health Science, Suzuka University of Medical Science, Suzuka, 510-0293, Japan. [email protected].
- Graduate School of Health Science, Suzuka University of Medical Science, Suzuka, 510-0293, Japan.
- Department of Electronic and Computer Engineering, Ritsumeikan University, Kusatsu, 525-8577, Japan.
- Department of Radiological Technology, Tsukuba Medical Center Hospital, Tsukuba, 305-8558, Japan.
Abstract
The uniform biennial mammography approach currently used in Japan's population-based breast cancer screening program has limitations. We developed a radiomics-based machine learning method with feature-level transparency to analyze image-derived patterns associated with future breast cancer occurrence using Japanese screening mammograms. The dataset comprised 208 mediolateral oblique mammograms obtained at the Tsukuba Comprehensive Health Screening Center, including 104 pre-symptomatic screening mammograms from 104 women subsequently diagnosed with breast cancer and 104 mammograms from 104 women with no abnormal findings. The pre-symptomatic mammograms were acquired 2-6 years before diagnosis and showed no definite suspicious findings at screening interpretation. For the future breast cancer group, 128 × 128-pixel regions of interest (ROIs) were extracted from lesion-corresponding areas on pre-symptomatic screening mammograms. For the healthy group, ROIs were extracted from the entire breast region. Following the Image Biomarker Standardization Initiative, 185 radiomics features were calculated per ROI. The least absolute shrinkage and selection operator selected 11 radiomics features. Using these features, five classifiers were compared: decision tree, k-nearest neighbors, naïve Bayes, linear support vector machine, and regularized discriminant analysis (RDA). In patient-level 5-fold cross-validation repeated five times using common fixed partitions, RDA achieved the highest mean ROI-based AUROC of 0.872 ± 0.003. After hyperparameter optimization using the same repeated cross-validation partitions, the final model achieved a breast-level AUROC of 0.803 (95% CI, 0.693-0.894) on the independent evaluation dataset. These findings suggest that local radiomic patterns in pre-symptomatic screening mammograms may contain discriminative information associated with future breast cancer occurrence in a case-control setting.