Spectrochimica acta. Part A, Molecular and biomolecular spectroscopy · 2026
Macromolecular characterization of basal cell carcinoma using ATR-FT-IR spectroscopy combined with machine learning.
СтатьяНазвание и аннотация: Перевод в очереди
Обработка статьи · этапы, ошибки и управление
Загрузка состояния…
Ошибки и ограничения
История · было → стало
Состояние обновляется каждые 5 секунд, пока блок открыт. Процент показывается только там, где известен объём работы. Пауза останавливает продолжение этапа; текущий запрос к источнику может завершиться.
Guleken Zozan, Uyulan Caglar, Yıldırım Hacer Ahsen, Demir Furkan, Makouei Sahar Taghi, Kılıc Güray, Sarıbal Devrim
Аннотация
BACKGROUND: Basal cell carcinoma (BCC) is the most common human malignancy, yet its diagnosis relies on invasive histopathological biopsy with considerable inter-observer variability. Attenuated Total Reflectance Fourier Transform Infrared (ATR-FT-IR) spectroscopy provides non-destructive macromolecular fingerprinting of tissues, and machine learning enables high-dimensional pattern recognition in spectral data. This study aimed to characterize macromolecular alterations in BCC tissues and develop optimized machine learning classifiers for BCC detection.
METHODS: ATR-FT-IR spectra were acquired from 56 speciments (28 BCC, 28 matched intra-individual control tissue). Three spectral regions were analyzed as independent datasets: F (800-1800 cm-1, 1039 features), L (900-1300 cm-1, 417 features), and M (2800-3000 cm-1, 210 features). Four feature selection strategies (SelectKBest, Recursive Feature Elimination [RFE], Random Forest Importances, Lasso L1 regularization) were combined with two ensemble classifiers (Random Forest, XGBoost) using an 80/20 stratified train-test split (random_state = 42). Performance was evaluated by test accuracy, ROC-AUC, sensitivity, specificity, and F1 score.
RESULTS: XGBoost consistently outperformed Random Forest across all spectral regions, with Random Forest exhibiting perfect training accuracy (100%) and high overfitting risk. In the F dataset (800-1800 cm-1), XGBoost with SelectKBest achieved the best result: test accuracy 0.75, ROC-AUC 0.79, sensitivity 0.80, specificity 0.71. In the L dataset (900-1300 cm-1), XGBoost with RFE achieved test accuracy 0.83, ROC-AUC 0.83. In the M dataset (2800-3000 cm-1), XGBoost with Lasso achieved the highest single-split performance, with test accuracy of 0.92, ROC-AUC of 1.00, and specificity of 1.00. However, repeated stratified cross-validation yielded attenuated but supportive estimates, confirming the lipid CH stretching region as the most discriminative spectral domain while indicating that the single-split ROC-AUC should be interpreted cautiously. Best single-split performance was achieved using XGBoost + Lasso on the lipid region (2800-3000 cm-1), while repeated cross-validation supported the M dataset as the strongest but not perfectly generalizable spectral domain.
CONCLUSION: ATR-FT-IR spectroscopy combined with XGBoost and Lasso feature selection demonstrates high discriminative capacity for BCC detection, particularly in the lipid CH₂/CH₃ stretching region. These findings support the translational potential of FT-IR-based spectral diagnostics as a minimally invasive adjunct to histopathological confirmation.
METHODS: ATR-FT-IR spectra were acquired from 56 speciments (28 BCC, 28 matched intra-individual control tissue). Three spectral regions were analyzed as independent datasets: F (800-1800 cm-1, 1039 features), L (900-1300 cm-1, 417 features), and M (2800-3000 cm-1, 210 features). Four feature selection strategies (SelectKBest, Recursive Feature Elimination [RFE], Random Forest Importances, Lasso L1 regularization) were combined with two ensemble classifiers (Random Forest, XGBoost) using an 80/20 stratified train-test split (random_state = 42). Performance was evaluated by test accuracy, ROC-AUC, sensitivity, specificity, and F1 score.
RESULTS: XGBoost consistently outperformed Random Forest across all spectral regions, with Random Forest exhibiting perfect training accuracy (100%) and high overfitting risk. In the F dataset (800-1800 cm-1), XGBoost with SelectKBest achieved the best result: test accuracy 0.75, ROC-AUC 0.79, sensitivity 0.80, specificity 0.71. In the L dataset (900-1300 cm-1), XGBoost with RFE achieved test accuracy 0.83, ROC-AUC 0.83. In the M dataset (2800-3000 cm-1), XGBoost with Lasso achieved the highest single-split performance, with test accuracy of 0.92, ROC-AUC of 1.00, and specificity of 1.00. However, repeated stratified cross-validation yielded attenuated but supportive estimates, confirming the lipid CH stretching region as the most discriminative spectral domain while indicating that the single-split ROC-AUC should be interpreted cautiously. Best single-split performance was achieved using XGBoost + Lasso on the lipid region (2800-3000 cm-1), while repeated cross-validation supported the M dataset as the strongest but not perfectly generalizable spectral domain.
CONCLUSION: ATR-FT-IR spectroscopy combined with XGBoost and Lasso feature selection demonstrates high discriminative capacity for BCC detection, particularly in the lipid CH₂/CH₃ stretching region. These findings support the translational potential of FT-IR-based spectral diagnostics as a minimally invasive adjunct to histopathological confirmation.