Micro-CT radiomics reveals the microcalcification-surrounding tissue interface as the diagnostic epicenter of breast microcalcifications.

backgroundIn clinical breast imaging, microcalcifications (MCs) are routinely interpreted as small radiological signs, typically analyzed in clusters. They guide radiologists in assessing breast lesions and determining the likelihood of malignancy. Yet, increasing biological evidence indicates that active processes occur not only within the calcified core but also in the microcalcification-surrounding tissue microenvironment (MCST). This raises the possibility that current assessments may overlook critical diagnostic information embedded in the MCST. However, conventional mammography and digital breast tomosynthesis (DBT) lack the spatial resolution required to determine where diagnostic information truly resides. MATERIALS AND

methodsWe used a high-resolution (≈8 μm) 3D micro-CT scanner to scan 94 paraffin-embedded breast biopsy blocks, from which 3504 individual MCs were segmented to obtain a binary 3D mask ([Formula: see text]) for each MC. Unlike previous studies analyzing clustered MCs at mammographic resolution, our analysis operates at the level of individual MCs. Radiomic features were extracted from each MC and used to train machine-learning classifiers to predict the histopathological label (benign or malignant) of the lesion in which each MC occurred. To determine (i) whether discriminative information arises predominantly from the calcified core or from the MCST and (ii) to separate genuine tissue signal from preprocessing or segmentation effects, we conducted three complementary analyses. First, we held [Formula: see text] constant and varied only the size of the rectangular preprocessing window around [Formula: see text], i.e. the 3D region of grayscale data considered for feature computation. Second, we assessed robustness to segmentation by making small, incremental changes to [Formula: see text] (erosions to restrict the calcified core; dilations to extend into immediately adjacent tissue). Third, to test MCST-only signal, we classified MCs using features computed exclusively from concentric shells defined at incremental offsets from the [Formula: see text]: outer shells (thin layers of MCST just outside [Formula: see text]) and inner shells (thin layers within the calcified core, inside [Formula: see text]).

resultsIncreasing the size of the preprocessing window around a constant [Formula: see text] improved classification performance (AUC ≈ 0.688 → 0.811), revealing strong contextual effects in radiomics feature extraction. Moderate mask dilations likewise improved performance (AUC ≈ 0.689 → 0.813), indicating that the diagnostic signal extends beyond the calcified core into the MCST. Remarkably, even when using only concentric shell features, classification performance remained high (AUC ≈ 0.81).

conclusionThis study provides the first imaging-based evidence - at micrometer scale and in 3D - pinpointing where the discriminative information of breast MCs arises: within the calcified core and/or the MCST? From a radiomics and technical standpoint, the preprocessing window matters: it should be controlled and reported, as it directly affects extracted features and can influence diagnostic performance. We find that MCST is the dominant source of discriminative information for classifying individual MCs as benign or malignant: classification using features computed exclusively from concentric shells performs on par with maximally dilated mask (encompassing both the core and MCST) but clearly exceeds calcified-core-only. These findings shift the interpretative focus of MC analysis from the calcified deposits themselves to their immediate microenvironment, suggesting that future biomarkers and quantitative features should target MCST.