The study used a curated dataset of 504 anonymized CTA scans retrospectively collected at UZ Brussel (Brussels, Belgium) between January 2014 and December 2023 from patients evaluated for DIEP flap breast reconstruction. The inclusion criterion was that a CTA was performed for preoperative DIEP flap planning. CTA scans acquired for other clinical indications were excluded. Scans were acquired with standardized parameters using a GE Medical Systems Discovery CT750 HD scanner (GE Healthcare, Chicago, Illinois, United States) to ensure consistency across the dataset. All patients were scanned using the same DIEP CTA protocol as outlined by Karunanithy et al.^[18]. From this dataset, 55 patients were randomly selected for manual volumetric annotation of the DIEP vascular anatomy. A train-test split strategy was used. Five scans were used for model hyperparameter tuning, and the remaining 50 were used for 5-fold cross-validation.

A separate cohort of 100 patients was randomly selected for manual annotation of the umbilical landmark location (x, y, z coordinates).

Each CTA volume underwent preprocessing, including conversion from digital imaging and communications in medicine (DICOM) to neuroimaging informatics technology initiative (NIfTI) format, intensity normalization and clipping to the range of 0-1, and resampling to an isotropic voxel size of 1 × 1 × 1 mm. Statistical analyses were performed using Python 3.11 (Python Software Foundation, Wilmington, Delaware, United States).

To overcome the challenges of annotating narrow, low-contrast perforator vessels in full-volume CTA, a novel dual-maximum intensity projection (MIP) annotation workflow was implemented [Figure 2]. The first major step was to back-project MIP labels into the CTA image domain. This approach mirrors established clinical practice, in which radiologists rely on slab-based MIPs for visualizing intricate vascular networks. Annotators first independently reviewed and annotated 10-mm-thick axial and coronal MIP slabs, which enhanced vessel conspicuity and continuity. A depth-map algorithm was then used to recover the precise 3D spatial coordinates lost during MIP creation. Annotations were performed in 3D Slicer (The Slicer Community)^[19], and these MIP-enhanced annotations were back-projected into the original 3D CTA volume.

To optimize computational efficiency and limit the analysis to clinically relevant anatomy, an automated region of interest (ROI) localization module was developed [Figure 3]. Processing full CTA volumes is both time- and memory-intensive because large portions of the volume are unrelated to perforator mapping.

The ROI localization was implemented using a two-step approach. First, atlas-free landmark-guided segmentation was performed using pretrained models from TotalSegmentator^[20]. The volume of interest was constrained using anatomically identifiable landmarks: the left and right anterior superior iliac spines (ASIS) defined the lateral boundaries, the pubic tubercle defined the caudal limit, and a point 5 cm inferior to the xiphisternum defined the cranial limit. The anterior boundary was set as the most anterior skin surface, with an additional 2-cm margin, whereas the posterior boundary corresponded to the plane of the bilateral ASIS. Second, the umbilicus was localized using a heatmap regression approach based on a U-Net architecture (n = 100). To detect the location of the umbilicus, a heatmap-regression 3D U-Net method was introduced that predicts the heatmap peak at the umbilicus location. The method used binary cross-entropy loss and a 256 × 256 × 256 voxel input patch for training. The aim was to establish the umbilicus as the anatomical reference point for future automated quantitative reporting of perforator location.

Within the localized ROI, a deep learning model was trained for perforator vessel segmentation. The 3D Swin UNETR (Swin Transformer-based) architecture and the Attention U-Net architectures were evaluated for their ability to capture long-range dependencies and global context, which are crucial for segmenting elongated and tortuous vascular networks^[21,22]. These models process 3D image patches and learn to identify and delineate perforator vessels.

The training strategy for the deep learning model involved:

1. Patch sampling: Patches of 96 × 96 × 96 voxels were extracted to ensure a balanced representation of vessel and non-vessel regions, enabling the network to learn contextual features from both vessel-rich and vessel-poor areas.

2. Loss function: A composite loss function combining Dice similarity coefficient (Dice) loss (for overall overlap) and cross-entropy loss (for pixel-level accuracy) was used, with a weighting that prioritized Dice loss to encourage correct overall region shape and overlap.

3. Data augmentation: A comprehensive set of 3D data augmentations (e.g., random flips, rotations, scaling, and noise injection) was applied to simulate inter-scanner and inter-patient variability, enhancing the model’s generalizability.

4. Optimization: The AdamW optimizer^[23] was used with a learning rate scheduler combining linear warmup and cosine annealing to stabilize and fine-tune training.

Post-processing steps were systematically applied to the raw model outputs to refine predictions and enhance topological consistency, both of which are critical for clinical interpretation. These steps included thresholding, removing small, isolated components, extracting centerlines, and bridging minor discontinuities, thereby ensuring continuous and anatomically plausible vessel segmentations.

The primary development environment was built using an NVIDIA Ampere GPU with 40 GB of memory. Software development and experimentation were performed in Python using PyTorch (Meta AI, New York City, United States) and the Medical Open Network for AI (MONAI) (NVIDIA, National Institutes of Health, and King’s College London) frameworks^[24].

To address limitations of the baseline model, particularly incomplete segmentation of small vessels and discontinuities in fine branching structures, two targeted enhancements were evaluated for sparse perforator networks.

1. Continuity-Aware loss function: To improve sensitivity to thin, tubular structures,

the baseline loss function was modified by adding the centerline Dice (clDice) loss component.

2. Multiclass segmentation: To bias the network toward underrepresented, thin perforators primarily located in the adipose region, a fat mask was used to first create multiclass ground-truth annotations and to increase patch sampling within the fat regions.

To assess how well the segmentation model performed in identifying and outlining the perforator vessels, we used a combination of standard and anatomy-specific evaluation metrics^[25] selected for their clinical relevance in flap planning:

• Dice: This metric measures how closely the segmented vessels match the reference anatomy in terms of overall shape and volume. A score of 1 indicates perfect overlap, whereas lower scores indicate discrepancies in vessel detection.

• clDice: This metric evaluates how well the model captures the course and branching pattern of the vessels by comparing the predicted centerlines or “skeletons” with the reference centerlines. It is especially useful for assessing whether fine branches are continuous and correctly traced, which is important for surgical planning, where vessel continuity matters.

• Recall: This metric indicates how many reference vessels were correctly identified by the model. Higher recall indicates fewer missed perforators, which is particularly important for small or deep branches in adipose tissue.

• 95th Percentile Hausdorff Distance (HD95): This metric measures the worst-case spatial error between the segmented vessel borders and the true anatomy while reducing the influence of extreme outliers. It provides insight into how precisely the vessel edges are mapped, which can affect flap design or intraoperative navigation.

Together, these outcome measures provide a balanced assessment of the model’s performance, including whether it accurately captures the key vessels, preserves their connectivity, and achieves a level of detail suitable for microsurgical decision-making. Statistical analysis was performed using Wilcoxon signed-rank tests with Bonferroni correction, with results considered significant at P < 0.05.

The depth-aware back-projection of MIP images substantially reduced annotation time by 60%-70% (from approximately 4 h to approximately 90 min per patient) and significantly improved vessel continuity while reducing the number of missed perforators. Additionally, this approach better aligned the annotation process with current clinical practice.

The atlas-free, segmentation-based ROI cropping strategy demonstrated excellent performance, successfully generating skeletal and sternum segmentations and automatically extracting all target anatomical landmarks in 100% of the evaluated scans. The process was robust, with a mean inference time of 28 ± 5 s per volume, supporting its integration into a rapid clinical workflow.

The U-Net + heatmap-based umbilicus detector achieved a mean target registration error (TRE) of approximately 2 mm in the validation cohort, enhancing the stability and reproducibility of ROI definition, particularly in cases with variable abdominal morphology or surgical scarring.

The baseline segmentation model, Swin UNETR, demonstrated robust performance for identifying the main perforator trunks and proximal secondary branches. It achieved a median global Dice coefficient of 0.58 [0.54, 0.63], significantly outperforming Attention U-Net (Dice 0.55 [0.49, 0.60], P < 0.05). Recall was also significantly higher for Swin UNETR than for Attention U-Net (0.54 [0.47, 0.60] versus 0.43 [0.35, 0.50], P < 0.05), reflecting superior sensitivity to actual vessel voxels. Boundary delineation was more consistent with a significantly lower HD95 of 20.0 mm [15.0, 23.8] versus 33.5 mm [25.0, 36.7] (P < 0.05), indicating closer alignment with the ground truth. Topological continuity, measured by clDice, was similar between the two models (0.41 for Swin UNETR versus 0.40 for Attention U-Net).

Visually, Swin UNETR reliably recovered the principal perforators and main branches but consistently missed smaller-caliber branches and produced small gaps at vessel bifurcations, particularly in deeper adipose tissue. In simpler anatomical cases, such as sparse perforator networks, performance was notably better, with Dice values reaching up to 0.72.

Based on this superior performance, Swin UNETR was adopted as the baseline model for subsequent experiments.

Incorporating a continuity-aware component, clDice, into the standard Dice and cross-entropy loss functions significantly improved model performance. Values are presented to two decimal places, with 95% confidence intervals (CIs) shown in brackets.

• Global Dice improved to 0.60 [0.54, 0.65] (P ≤ 0.05), compared to a baseline (experiment-specific control) Dice of 0.58 [0.54, 0.63].

• Recall increased significantly to 0.58 [0.51, 0.64] (P ≤ 0.05), reflecting improved sensitivity and fewer missed vessels.

• HD95 decreased notably to 17.5 mm [11.1, 23.4], highlighting tighter and more accurate vessel boundary delineation.

• clDice remained stable at 0.41 [0.38, 0.46], indicating maintained or slightly improved vessel continuity.

• This enhancement produced smoother and more anatomically coherent vessel segmentations, reducing the fragmentation of vessel branches critical for preoperative perforator mapping.

Multiclass segmentation also significantly improved segmentation outcomes by leveraging a pretrained fat and muscle segmentation model to guide region-aware learning^[26]:

• Global Dice increased to 0.60 [0.56, 0.63], P ≤ 0.05 from the experiment-specific control baseline of 0.58, matching the continuity-aware method’s overall gain.

• Notably, Dice in the challenging adipose compartment increased to 0.52 [0.47, 0.57] compared with the baseline value of 0.49 [0.40, 0.55], representing a clinically meaningful improvement in the region most relevant to DIEP flap harvesting.

• HD95 decreased slightly to 19.15 mm [12.2, 24.7], indicating better boundary precision in adipose areas.

• clDice showed moderate improvement to approximately 0.41 [0.35, 0.44] (P ≤ 0.05), suggesting better continuity, although distal vessel tips remained partially under-segmented.

• From a clinical perspective, this method enhanced visibility and segmentation quality precisely in the critical fat regions used for surgical planning, without compromising global segmentation quality.

Figure 4 compares two representative cases: complex perforator anatomy (Example 1) and sparse perforators (Example 2). For each case, the ground truth, baseline, and both enhanced model outputs are shown to illustrate differences in segmentation predictions.