Abstract
BACKGROUND: Early diagnosis of breast cancer (BC) is crucial for improving patient outcomes. Features of the peritumoral region have been shown to significantly enhance the predictive performance of deep learning radiomics (DLR) models. This study aims to explore the impact of key parameter selection on improving the performance of the intratumoral-peritumoral region fusion model. The goal is to enhance the modal's non-invasive diagnostic capability for distinguishing benign and malignant breast tumors. MATERIALS AND METHODS: This retrospective study included 411 female patients with breast lesions from four hospitals. DLR models were constructed using their contrast-enhanced ultrasound (CEUS) images. The intratumoral region of interest (ROI) was gradually expanded to generate peritumoral regions of varying thicknesses. Six groups of fusion models were constructed using different key parameter combinations, including pseudo-color (PC) vs. grayscale (GRAY) images, original precise (OP) ROI vs. bounding box (BB) ROI, and direct extension (DE) strategy vs. feature-level fusion (FLF) strategy. Additionally, a reader study was conducted, comparing the diagnostic performance of the best fusion model with that of six radiologists. The performance of the models was evaluated using the area under the curve (AUC). RESULTS: Incorporating the peritumoral region significantly enhanced the diagnostic performance of the DLR models. The PC-OP-DE-Peri (4mm) model achieved the highest performance in the testing cohort, with an AUC of 0.837. The performance surpassed both the intratumoral models and all radiologists. The effects of different key parameter selections on fusion model performance varied. CONCLUSION: This study suggests that the selection of PC images, OP ROIs, and the DE strategy effectively improves the performance of intratumoral-peritumoral region fusion models for predicting BC.