Optimising neural networks for perforator detection in DIEP flap breast reconstruction using dynamic infrared thermography

Breast reconstruction following mastectomy is increasingly performed, with Deep Inferior Epigastric artery Perforator (DIEP) flap surgery considered the gold standard. Accurate preoperative perforator selection is vital to minimize complications and operative time. While computed tomography angiography (CTA) remains the clinical reference, drawbacks including radiation, contrast use, and cost motivate exploration of non-invasive alternatives. Dynamic Infrared Thermography (DIRT) offers a low-cost, radiation-free method but still lacks automation. This study evaluates deep learning for automated perforator detection in DIRT. A dataset of 50 time-lapse thermograms from five patients was acquired using various cooling methods and validated through leave-one-out cross-validation (LOOCV). Two neural network architectures were compared: a standard U-Net and a modified U-Net (mU-Net) from prior work. U-Net consistently outperformed mU-Net. Across LOOCV folds, U-Net achieved a mean weighted Dice loss of 0.42 ± 0.09, sensitivity of 0.87 ± 0.08, and precision of 0.82 ± 0.14. On an independent test patient, sensitivity remained high (0.88) but precision decreased (0.58). The mU-Net failed to converge (validation loss 0.87 ± 0.02), producing uniform segmentations. These findings demonstrate that U-Net is a robust tool for automated perforator detection in DIRT, though false positives highlight the need for larger datasets and further optimisation before clinical use.