The study focuses on breast reconstruction planning using a DIEAP flap, where manual identification of perforating vessels from CT angiography is laborious and variable. They developed a novel AI pipeline that extracts anatomical vessel axes from CTA data and uses them to drive deep learning segmentation models. They compared the zero-shot models SAM 2, MedSAM-2 and nnInteractive, with the best-performing nnInteractive fine-tuning clinical data with Recall Loss to preserve vessel topology. On the test set of 9 patients, the cube similarity coefficient increased to 0.2755 from the original 0.4. The tuned model produced more anatomically acceptable and continuous vessel segmentations. The pipeline automatically quantified metrics such as the length of the intramuscular path of the perforators and their distance from the umbilicus. This AI tool reduces manual annotation and improves the consistency of preoperative planning.