| Title |
Supervised machine learning thyroid carcinoma diagnosis using wide-field SHG microscopy |
| Authors |
Padrez, Yaraslau ; Hristu, Radu ; Timoshchenko, Igor ; Eftimie, Lucian G ; Rutkauskas, Danielis ; Golubewa, Lena |
| DOI |
10.1109/ACCESS.2025.3583435 |
| Full Text |
|
| Is Part of |
IEEE access.. Piscataway, NJ : IEEE. 2025, vol. 13, p. 112021-112038.. ISSN 2169-3536 |
| Keywords [eng] |
association rule learning ; automated machine learning ; boosting ; biomedical image processing ; cancer ; image analysis ; wide-field second harmonic generation microscopy |
| Abstract [eng] |
Papillary (PTC) and follicular (FTC) thyroid carcinomas require different treatment strategies, but their accurate differentiation remains a challenge in conventional histopathology. Misclassification can lead to overtreatment of low-risk PTCs or inadequate treatment of FTCs, increasing the risk of recurrence and metastasis. Since the structure of the collagen capsule surrounding thyroid nodules provides diagnostically valuable information, label-free imaging with second harmonic generation (SHG) microscopy combined with machine learning (ML)-based analysis offers a promising approach for automated classification. In this study, we extracted intensity and texture features from SHG images of thyroid nodules scanned as a whole and optimized several ML classifiers, including logistic regression, support vector classification (C-SVC), multilayer perceptron, random forest, XGBoost, and LightGBM, using hyperparameters tuning with stratified 10-fold cross-validation. One of the major challenges in classification was label noise resulting from 1) mislabeling of adjacent tissue, 2) PTC calcifications mimicking FTC features, and 3) capsule heterogeneity. To address this issue, we applied unsupervised segmentation to exclude mislabeled regions and consider capsular heterogeneity as a diagnostic feature. Recursive feature elimination and mutual information selection further refined the feature set and improved classification accuracy. Among all models, C-SVC achieved the highest accuracy (84.73%) with robust generalization to unknown data, significantly outperforming standard ML approaches (60–70%). These results demonstrate the feasibility of SHG microscopy-based ML classification as a reliable adjunct to existing histopathologic methods, which could improve diagnostic accuracy and patient outcomes. |
| Published |
Piscataway, NJ : IEEE |
| Type |
Journal article |
| Language |
English |
| Publication date |
2025 |
| CC license |
|
