Machine learning assisted mid-infrared spectrochemical fibrillar collagen imaging in clinical tissues

Label-free multimodal imaging methods are essential for comprehensive tissue analysis, particularly for studying the tumor microenvironment where changes in fibrillar collagen architecture are linked to cancer progression. This study presents a novel multimodal computational imaging approach combining mid-infrared spectral imaging (MIRSI) with second harmonic generation (SHG) microscopy to detect fibrillar collagen in biological tissues. Using SHG images as ground truth, a supervised machine learning (ML) model, specifically a random forest (RF) classifier, was trained to classify fibrillar collagen in mid-infrared hyperspectral images. The study utilized five human pancreatic tissue samples, encompassing 2.8 million MIRSI spectra for training and 68 million spectra for validation. The RF-MIRSI model demonstrated robust performance in collagen segmentation (boundary F-score: 0.8 at 4-pixel thresholds), orientation (Pearson’s R: 0.82), and alignment (Pearson’s R: 0.66) when compared to SHG ground truth. These findings highlight the potential of ML-driven label-free MIRSI for detailed collagen fiber and tumor microenvironment analysis, offering a complementary structural and chemical imaging tool for tumor pathology.