Related Experiment Video
Updated: Mar 31, 2026
07:57
An Experimental Human DIEP Flap Model to Investigate Preservation Strategies for Vascularized Composite Allografts and Free Flaps
Published on: December 5, 2025
403
AI-based planning for DIEAP flap procedures: exploring foundation models for artery perforators analysis.
Matilde Andrade1,2, Nuno Loução1, David Pinto1,3
1Digital Surgery Lab, Breast Cancer Research Program, Champalimaud Foundation, Lisbon, Portugal.
Frontiers in Medicine
|March 30, 2026
Summary
This study introduces an AI pipeline for automated perforator vessel mapping in Deep Inferior Epigastric Artery Perforator (DIEAP) flap surgery. The AI model significantly improves the accuracy and efficiency of preoperative planning, enhancing surgical consistency.
Area of Science:
- Medical Imaging
- Artificial Intelligence in Surgery
- Computational Anatomy
Background:
- Deep Inferior Epigastric Artery Perforator (DIEAP) flap breast reconstruction is a standard autologous procedure.
- Preoperative planning, particularly identifying perforator vessels from Computed Tomography Angiography (CTA) images, is manual, time-consuming, and prone to variability.
- Enhancing the efficiency and consistency of this planning process is crucial for successful surgical outcomes.
Purpose of the Study:
- To develop and validate an automated, end-to-end AI pipeline for segmenting and quantitatively analyzing perforator vessels in CTA images.
- To improve the efficiency and consistency of preoperative planning for DIEAP flap breast reconstruction.
- To assess the performance of foundation models in a zero-shot and fine-tuned setting for this specific application.
Main Methods:
- A novel pipeline was developed using computer vision for initial vessel centerline extraction from CTA data.
- These centerlines guided a Deep Learning (DL) segmentation model, with three foundation models (SAM 2, MedSAM-2, nnInteractive) benchmarked.
- The best-performing model, nnInteractive, was fine-tuned using a connectivity-aware compound loss with Skeleton Recall Loss (SRL) to preserve vessel topology.
Main Results:
- The fine-tuned nnInteractive model achieved a mean Dice Similarity Coefficient (DSC) of 0.265, a significant improvement from the 0.174 zero-shot baseline.
- Qualitative assessment showed more anatomically plausible and continuous vessel segmentations compared to the baseline.
- The automated pipeline successfully quantified key surgical planning metrics, including intramuscular path length and distance to the umbilicus.
Conclusions:
- An end-to-end, AI-driven workflow for perforator mapping in DIEAP flap planning is feasible.
- The use of foundation models with anatomical priors and topology-aware fine-tuning reduces manual annotation burden and improves consistency.
- This automated pipeline offers a promising approach to enhance preoperative planning efficiency and reliability, potentially improving surgical outcomes.
