卵巣がん細胞の分類を支援する機械学習支援ラマンスペクトル解析

Ovarian cancer is one of the most lethal gynecological malignancies, asymptomatic early progression, ineffective screening, and high histological heterogeneity. Accurate subtype classification and detection of chemotherapy resistance are critical for guiding personalized treatment strategies. Raman spectroscopy offers a label-free, non-destructive means of capturing biochemical fingerprints of cells, but its clinical potential is hindered by high spectral complexity and subtle inter-class variations. This study presents a machine learning-assisted Raman spectroscopy framework for the classification of ovarian cancer cell subtypes and their cisplatin resistance phenotypes. Raman spectra were acquired from normal ovarian epithelial cells (IOSE-80), four ovarian cancer cell lines (A2780, SKOV3, OVCAR-3, ES-2), and cisplatin-resistant variants (A2780-DDP, SKOV3-DDP). Three computational models were developed and systematically compared: a principal component analysis-support vector machine (PCA-SVM) algorithm and two convolutional neural network (CNN-Enhance and CNN-BiLSTM). Classification performance was assessed across three tasks: (i) discrimination of normal versus malignant cells, (ii) differentiation of cancer cells from their cisplatin-resistant variants, and (iii) classification of distinct cancer subtypes. Results show that Raman spectra reveal distinctive biochemical differences between normal and malignant cells, particularly in protein-, lipid-, and nucleic acid-related peaks. Both PCA-SVM and CNN achieved high classification accuracy (>90%) in most tasks, with PCA-SVM demonstrating greater stability and superior performance in subtype classification, while CNN showed advantages in specific cell-type detection. Notably, PCA-SVM achieved up to 100% accuracy in differentiating cisplatin-resistant phenotypes. These findings demonstrate that integrating Raman spectroscopy with machine learning enables label-free, and accurate classification of ovarian cancer subtypes and drug resistance, offering a promising pathway toward minimally invasive precision diagnostics and personalized cancer treatment planning.