Beyond the Tumor: Multi-region Radiomic and Deep Feature Analysis of Tumor, Peritumoral, and Posterior Imaging Features from Mammography and Ultrasound in Differentiating Invasive Micropapillary Carcinoma from Invasive Carcinoma of No Special Type

Background: Invasive micropapillary carcinoma (IMPC) is a rare breast cancer subtype with disproportionately aggressive biological behavior and a strong propensity for lymphovascular invasion. Conventional mammography and ultrasound often fail to clearly distinguish IMPC from invasive ductal carcinoma of no special type (IDC-NST), limiting preoperative risk stratification. Quantitative imaging—through radiomics and deep feature analysis—may reveal subtle tumor and microenvironmental signatures inaccessible to visual assessment. Methods: We retrospectively analyzed 8 pathologically confirmed IMPC cases and 24 stage- and age-matched IDC-NST controls. Tumor, peritumoral (2, 5, and 10 mm), and posterior acoustic regions (5, 10, and 20 mm) were manually annotated and algorithmically generated from mammography and ultrasound images. Macroscopic features (boundary type, intensity, entropy, echo characteristics) and 512-dimensional deep features extracted using a pretrained ResNet-18 model were compared between groups. Statistical testing included Wilcoxon rank-sum tests and Cohen’s d , with joint significance scoring integrating p -values and effect size. Results: Macroscopic differences between IMPC and IDC-NST were subtle. IMPC lesions demonstrated smoother boundaries on both ultrasound and mammography and reduced posterior acoustic heterogeneity at 10 mm depth (Entropy_mean: 5.47 ± 0.58 vs. 6.07 ± 0.52, p  = 0.023). In contrast, deep learning features showed pronounced modality- and region-specific separation. Mammography exhibited dense clusters of significant deep features, particularly within tumor and 2–5 mm peritumoral regions, while ultrasound revealed more modest significance concentrated at narrow peritumoral distances. Overall, 65% of strictly significant features arose from mammography-derived representations. Conclusion: Although macroscopic imaging findings offer limited discriminatory ability, multiregional deep features from mammography and ultrasound capture meaningful morphological and microenvironmental differences between IMPC and IDC-NST. These results highlight the potential of integrated multi-modality deep feature analysis to enhance preoperative recognition of IMPC and support future development of automated diagnostic and risk-stratification tools.