Slow Diffusion Coefficient Derived from High b -Value Diffusion-Weighted MRI for Assessing Pathological Complete Response after Neoadjuvant Chemotherapy in Breast Cancer: An Exploratory Machine Learning Analysis

Abstract

Background The apparent diffusion coefficient (ADC) is used to assess breast cancer response but may be influenced by perfusion and T2-related effects. Aim The aim of the study was to compare ADC-only, slow diffusion coefficient (SDC)-only, and combined ADC–SDC models and to evaluate whether SDC provides information complementary to ADC for post-treatment assessment of pathological complete response (pCR). Materials and methods This retrospective secondary analysis included 84 patients from the ACRIN-6698/I-SPY2 dataset, including 32 with pCR. ADC was calculated from b  = 0 and 800 s/mm 2 and SDC from b  = 600 and 800 s/mm 2 . Lesion-mean values and treatment-related changes were evaluated using 100 repetitions of nested fivefold cross-validation. All normalization, random-forest ranking, and LASSO tuning and selection were performed within the training data. Results The combined rcADC–rcSDC model achieved the highest mean cross-validated AUC of 0.724. Its aggregated out-of-fold AUC was 0.728, compared with 0.633 for rcADC alone; however, the paired AUC difference was not statistically significant. Conclusion Treatment-related SDC changes may complement conventional ADC changes in post-treatment pCR assessment. Although the combined rcADC–rcSDC model achieved the highest observed discrimination, a statistically significant improvement over that of rcADC alone was not demonstrated. These exploratory findings require technical and external validation.