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

Significance: Label-free multimodal imaging methods that can provide complementary structural and chemical information from the same sample are critical for comprehensive tissue analyses. These methods are specifically needed to study the complex tumor-microenvironment where fibrillar collagen's architectural changes are associated with cancer progression. To address this need, we present a multimodal computational imaging method where mid-infrared spectral imaging (MIRSI) is employed with second harmonic generation (SHG) microscopy to identify fibrillar collagen in biological tissues.

Aim: To demonstrate a multimodal approach where a morphology-specific contrast mechanism guides an MIRSI method to detect fibrillar collagen based on its chemical signatures.

Approach: We trained a supervised machine learning (ML) model using SHG images as ground truth collagen labels to classify fibrillar collagen in biological tissues based on their mid-infrared hyperspectral images. Five human pancreatic tissue samples (sizes are in the order of millimeters) were imaged by both MIRSI and SHG microscopes. In total, 2.8 million MIRSI spectra were used to train a random forest (RF) model. The other 68 million spectra were used to validate the collagen images generated by the RF-MIRSI model in terms of collagen segmentation, orientation, and alignment.

Results: Compared with the SHG ground truth, the generated RF-MIRSI collagen images achieved a high average boundary $F$ -score (0.8 at 4-pixel thresholds) in the collagen distribution, high correlation (Pearson's $R$ 0.82) in the collagen orientation, and similarly high correlation (Pearson's $R$ 0.66) in the collagen alignment.

Conclusions: We showed the potential of ML-aided label-free mid-infrared hyperspectral imaging for collagen fiber and tumor microenvironment analysis in tumor pathology samples.