Study population: We retrospectively analyzed patients who underwent UF-DCE MRI at our hospital from 2018 to 2021, diagnosed as or suspected of breast cancer. Those with post neoadjuvant chemotherapy and/or radiation therapy, no pathological diagnosis, benign lesions,poor visualization of tumor or vessels, and those with poor image quality were excluded. Bilateral lesions were regarded as 2 different lesions.
MRI acquisition: 3T-MRI (Prisma, Siemens, Erlangen) with a 18-ch breast coil was used. The UF-DCE protocol used a 3D gradient-echo VIBE with compressed sensing (Prototype sequence). Gadobutrol was injected at 2 ml/s. UF-DCE MRI (from 15 sec before to 60 sec after contrast injection, 2 sec preparation time followed by 3.65 sec/phase x 20 phases) were obtained.
Image analysis: Axial subtracted MIP images (2nd -20th phases) were computed and TRVs were automatically segmented using an in-house developed segmentation programme. The programme utilized a U-Net-based model which achieved a Dice coefficient of 0.82[5]. 
TRV length was defined as the total pixel count of the “skeletonized” vessels, in order to minimize the effect of vessel thickness (Vessel thickness in segmented TRV may be affected by image quality of UF-DCE MRI and might introduce bias) .The skeletonization was performed using Lee's method[3].
Among many phases of the images, the 18th phase was used to quantify vessels due to the superior visibility of the early-enhanced vesselsStandardized TRV length (sTRV length) was defined as TRV length/(diameter of the breast lesion in mm). TRV and sTRV length was compared between large (diameter >=20mm) lesions and small lesions, and among subtypes. Wilcoxon rank-sum test was used for two-group comparison.