A deep learning framework for histopathological analysis of pixel-level extracellular matrix variation in standard H&E-stained images.

Artificial intelligence-driven image analysis has enabled significant advances in digital pathology. However, most approaches have focused on cell or organ structures. This manuscript presents a reproducible deep learning methodology for pixel-level analysis of amorphic patterns in haematoxylin and eosin-stained whole-slide histological images. This study analysed the pixel patterns in the extracellular matrix (ECM) part of connective tissue to identify differences in airway wall ECM compartments and their heterogeneity, which are microscopically similar and difficult to discern with the human eye. Through a targeted preprocessing pipeline, the deep learning model is guided to emphasise learning from pixel-level patterns in non-cellular tissue components while reducing the influence of cellular structures and artefacts. Combined with transfer learning, the model accurately distinguishes the characteristics of the airway submucosa and adventitia, achieving a test area under the curve of 0.84. Using visualisation techniques and statistical analysis, we demonstrate that random pixel imputation successfully reduces the effects of cellular structures on model learning. The framework is applied in a proof-of-principle study of lung tissue from patients with chronic obstructive pulmonary disease, illustrating how this quantitative approach can study population heterogeneity and inform novel research directions. Ultimately, this study provides an innovative and adaptable framework that unlocks the analytical potential of often-overlooked amorphic components in AI-empowered histopathology.