*Purpose: Predicting the risk of developing breast cancer using mammogram images has shown promising effects when using deep learning-based models. Existing models however are mostly limited to the use of a single mammogram exam, while in reality, many women have multiple longitudinal prior screening exams that may span for a long range of many years, with varying temporal intervals among exams. The goal of this study is to develop a risk model to account for long-range priors with irregular inter-exam intervals.
*Methods and Materials: Two retrospective datasets were used: Cohort A (n = 590, 293 biopsy-proven breast cancer and 297 normal controls, max 8 priors, inter-exam interval: 12 to 24 months) from a mammogram screening cohort of a single institution; the CSAW-CC (Karolinska Case-Control) dataset (n = 6459, 406 biopsy-proven breast cancer and 6,053 normal controls, max 4 priors, inter-exam interval: 12 to 36 months). We proposed a Spatio-Temporal Mamba(STM) risk model that (1) handles in theory an unlimited number of priors, (2) encodes inter-exam intervals directly in the Mamba structure, and (3) uses temporal contrastive learning to align representations across irregular temporal intervals of exams. CC and MLO views were combined. Patient-wise five-fold cross validation was conducted to report (format: average ±std) C-index and time-specific AUCs of from 1- to 5-year risk prediction. Model performance was assessed using up to 4 and 8 priors, with comparisons to the Mirai model (using a single exam).
*Results: STM model on Cohort A (4 priors): C-index 0.737±0.02; 1-/2-/3-/4-/5-yr AUC 0.734/0.729/0.709/0.687/0.665 ± 0.02-0.03. STM on Cohort A (8 priors): C-index 0.739±0.02; AUC 0.748/0.745/0.733/0.709/0.682 ± 0.02-0.04. STM on CSAW-CC (4 priors): C-index 0.737±0.02 and AUC 0.734/0.729/0.709/0.687/0.665 ± 0.02-0.03. Mirai on Cohort A: C-index 0.685±0.02 and AUC 0.685/0.660/0.670/0.655/0.653 ± 0.02-0.03. Mirai on CSAW-CC: C-index 0.687±0.02 and AUC 0.705/0.680/0.664/0.664/0.630 ± 0.02-0.03.
*Conclusions: The STM model shows substantially improved performance when comparing to state-of-the-art risk model using a single exam. Using more priors appears to be more predictive.
*Clinical Relevance/Application: STM model can handle and benefit from multiple prior exams with irregular inter-exam intervals of a long-range screening history to more accurately predict breast cancer risk in real-world scenarios.

Margarita Louise Zuley MD

Appearances