First-order Analysis of Stromal Collagen in Chronic Pancreatitis using Cross- Modality Image Translation with Convolutional Neural Networks

Chronic pancreatitis (CP) is a pancreatic disorder that involves persistent inflammation. The extracellular environment is primarily composed of fibrillar collagen with other matrix components. Collagen is excessively deposited during CP which leads to fibrosis, a hallmark of the condition. Second harmonic generation (SHG) microscopy allows for label-free, high-resolution and precise visualization of collagen structures and fibrotic changes at a microscopic level allowing accurate and quantitative assessment of the tissue microarchitecture. However, the need of specialized equipment and high costs associated with SHG microscopy limit its widespread use. This study aims to explore an image-to-image translation method using a convolutional neural network (CNN) architecture to generate SHG-equivalent ‘collagen images’ from H&E-stained brightfield (BF) microscopy images. The collagen images were generated through the image-to-image translation approach and the similarity of the translated images with SHG microscopy images was validated using multiple methods. The pixel intensity-based histograms of these translated collagen images were analysed using first-order texture analysis, revealing significant alterations in collagen texture associated with CP. Machine learning (ML) based models were utilized to assess and compare the classification performance of models. The performance of the collagen features for the optimal classification model was then evaluated through ROC-AUC (receiver operating characteristics – area under the curve) analysis. The proposed image-to-image translational approach can potentially provide a detailed and quantitative assessment of stromal collagen, improving the understanding of disease mechanisms and may further complement traditional histopathological workflow.