Supervised Learning for Collagen Segmentation in Bright-Field Histology: A Comparative Evaluation of U-Net and MLP

Quantifying collagen in histological slides is essential for diagnosing and monitoring fibrosis. However, the combination of PicroSirius Red staining with polarized light microscopy, one of the standard techniques, requires expensive equipment and time-consuming procedures. This study investigated whether supervised machine learning algorithms applied to bright-field images could provide an automated and accurate method for collagen segmentation, reducing the dependence on polarized imaging.

A total of 140 histological images of renal, cardiac, and tendinous tissues from mice and rats stained with PicroSirius Red were analyzed. The images were fragmented into smaller patches, and the green channel was selected for non-polarized analysis. The reference standard was generated from polarized images binarized using ImageJ. Two supervised models were evaluated: (1) a Multilayer Perceptron (MLP) trained using five features (RGB intensities, median filter, and Wavelet), and (2) a U-Net architecture adapted to the image dimensions.

Both models performed well, with the best results observed in rat tendon samples (Dice and precision > 80%). Performance was lower in mouse kidney images, possibly due to weaker staining and thinner collagen fibers. Overall, the MLP slightly outperformed the U-Net, which still showed comparable results.

The proposed methodology proved effective for collagen segmentation highlighting the potential of supervised learning to reduce costs and processing time in histological analysis. Future improvements include expanding the dataset and refining the reference standard to enhance robustness in biological tissue evaluation.