Toward Modality-Aware Preprocessing for Deep Learning-Based Blood Vessel Segmentation

Background: Preprocessing pipelines developed for retinal fundus photography are routinely transferred to other vascular imaging modalities without validation. Methods: Seven preprocessing stages from the retinal literature were assembled into a rule-based chain gated by image-quality descriptors calibrated from training-set percentiles. A ten-condition ablation (the complete chain, seven chains each omitting one stage, and two minimal chains of two stages each) with five-fold cross-validation was run on VessMAP, 100 confocal images of mouse cortical vasculature, and six conditions were repeated on the DRIVE and CHASE_DB1 fundus datasets. Results: On VessMAP, only removing intensity normalisation significantly reduced Dice (−0.017, paired Wilcoxon p<10−4); removing local contrast normalisation raised it by 0.019, and two-stage chains exceeded the complete chain by 0.025–0.027. Removing local contrast normalisation or denoising, or reducing the chain to two stages, also improved Dice on both fundus datasets. With thresholds fitted on training folds only, VessMAP Dice was 0.882±0.006, within the inter-annotator interval [0.864, 0.891], and a 59,034-parameter network came within 0.012 Dice of a 20.9-million-parameter model. Conclusions: The complete chain was not the best configuration on any dataset, including the fundus datasets its stages came from; preprocessing should be validated stage by stage on the target data.