AI-based planning for DIEAP flap procedures: exploring foundation models for artery perforators analysis

Breast reconstruction using the Deep Inferior Epigastric Artery Perforator (DIEAP) flap is the gold standard for autologous procedures, but its success relies on challenging preoperative planning. Identifying perforator vessels from Computed Tomography Angiography (CTA) images is currently a manual, labor-intensive, and variable process. This study's objective was to assess, fine-tune and validate an automated, end-to-end model-driven pipeline for the segmentation and quantitative analysis of perforator vessels to enhance planning efficiency and consistency.

We developed a novel pipeline that first uses computer vision algorithms to extract anatomical priors and generate initial vessel centerlines from CTA data. These centerlines were then used as spatial prompts to guide a Deep Learning (DL) segmentation model. We benchmarked three state-of-the-art foundation models (SAM 2, MedSAM-2, and nnInteractive) in a zero-shot setting. The best-performing model, nnInteractive, was subsequently fine-tuned on our clinical dataset using a connectivity-aware compound loss incorporating Skeleton Recall Loss (SRL) to preserve vessel topology.

The fine-tuned nnInteractive model demonstrated significantly improved performance on a held-out test set of nine patients, increasing the mean Dice Similarity Coefficient (DSC) from a 0.174 zero-shot baseline to 0.265. Qualitatively, the fine-tuned model produced more anatomically plausible and continuous vessel segmentations compared to the baseline. Furthermore, the automated pipeline successfully quantified critical surgical planning metrics from the segmentations, including the perforators' intramuscular path length and their distance to the umbilicus.

This study demonstrates the feasibility of an end-to-end, artificial intelligence (AI)-driven workflow for perforator mapping in DIEAP flap planning. The use of foundation models guided by anatomical priors and enhanced with topology-aware fine-tuning establishes a robust method for reducing manual annotation burden and improving consistency. This automated pipeline is a promising tool to support more efficient and reliable preoperative planning, ultimately poised to improve surgical outcomes.