Virtual Collagen Mapping for Quantitative Microarchitectural Analysis in Whole-Core Pancreatic Tissue Microarrays

Collagen remodelling plays a critical role in the progression of pancreatic intraepithelial neoplasia (PanIN) to pancreatic ductal adenocarcinoma (PDAC), though misdiagnosis often occurs due to the resemblance with chronic pancreatitis (CP). While second harmonic generation (SHG) microscopy is the current gold standard in research for visualizing fibrillar collagen, it is not utilized for routine histopathological use. This study highlights the need for an alternative method for quantitative collagen analysis that is compatible with standard hematoxylin and-eosin (HE) stained histopathological slides. Using the computational image-to-image translation approach, whole-core standard brightfield HE images of pancreatic tissues were translated into the new collagen images. From these computationally translated collagen images, quantitative collagen measures were extracted for PDAC, PanIN, CP, and normal pancreatic tissues, and their statistical significance was evaluated. Among the four classes, PDAC tissues exhibited the highest collagen alignment (n = 58, p < 0.01, R² = 0.2594), while normal tissues showed the lowest fiber density (n = 58, p < 0.0001, R² = 0.3569). Seven machine learning models were assessed to differentiate neoplastic from non-neoplastic tissues cores based on collagen measures extracted from the translated collagen images. Feature importance and ROC-AUC analyses both identified fiber length and collagen deposition area as the most prominent parameters, achieving AUC values of 0.83 and 0.73, respectively, for distinguishing between the tissue classes. The findings demonstrate that the collagen images derived from standard HE stained tissue microarrays (TMAs) whole-core images enable quantitative assessment of collagen remodelling across pancreatic lesions. Additionally, the study indicates that cross-modality image translation, as a cost-effective alternative, has the potential to offer deeper insights into histopathology and tissue microenvironment analysis without requiring collagen-specific imaging equipment or external labels, thereby further supporting its feasibility for integration into standard clinical workflows.