| Abstract [eng] |
This work investigates the integration of wide-field Second Harmonic Generation (SHG) microscopy and machine learning (ML) algorithms for the diagnosis of thyroid cancer and fibrosis in cancer-related diseases. SHG microscopy, a label-free imaging technique, was used to image collagen network structures in lung tissue affected by pulmonary arterial hypertension (PAH) and thyroid carcinoma nodules. Statistical and textural analysis of SHG images revealed stage-specific restructuring of collagen in PAH, demonstrating its potential as a non-destructive tool to assess fibrosis in PAH and related lung diseases. In thyroid pathology, polarization-resolved SHG combined with unsupervised ML enabled quantitative characterization of collagen capsule heterogeneity and collagen ultrastructure in papillary thyroid carcinomas and identified intact, invaded and microinvasive regions often overlooked in routine histology. Furthermore, supervised ML models, in particular the C-support vector classifier and multilayer perceptron (deep learning model), were able to effectively discriminate between follicular thyroid carcinomas and papillary thyroid carcinomas despite high data noise and achieved robust diagnostic accuracy. The study highlights SHG microscopy augmented with ML as an efficient complementary diagnostic tool to improve accuracy, reduce observer variability and support the development of automated classification systems. |
