A reproducible deep learning pipeline for collagen quantification in cardiac histology, benchmarked against colour deconvolution and inter-observer agreement

Abstract

Collagen area fraction is a standard histological measure of cardiac interstitial fibrosis, but is usually measured using colour thresholding and requires close image-by-image monitoring, which depends on the operator and cannot separate collagen from similarly coloured muscle. We describe an open, reproducible deep learning pipeline for collagen segmentation in mouse heart sections and test it against a calibrated colour-deconvolution baseline and human annotators. Seventy Masson’s trichrome fields from 29 mice were traced under a written protocol and partitioned by animal. On eight mice withheld from training and validation, six segmentation architectures reached a mean Dice of 0.787 to 0.826 against 0.397 for the baseline, which detected most collagen but also included stained muscle. Predicted area fraction tracked the manual reference closely (correlation 0.960 to 0.989). Mean pairwise Dice between three annotators was 0.846. Against two-observer consensus masks, U-Net ++ achieved a mean Dice of 0.795 compared with 0.829 for the observers; this exploratory comparison does not establish equivalence. Three topology-aware objectives, one proposed here, lowered Dice in every test animal. A self-configuring nnU-Net, without hyperparameter search, gave the highest Dice (0.838 over three runs) but did not exceed the agreement between the reference annotator and the closest second annotator (0.841). On picrosirius red, trichrome U-Net ++ labelled pale background instead of red-stained collagen, while one retrained on that stain reached 0.855 on perivascular and 0.704 on interstitial fields. The pipeline provides a reproducible approach to cardiac collagen quantification, while the observer study highlights annotation variability as an important consideration in performance assessment.