This study evaluated the diagnostic potential of Fourier-transform infrared (FTIR) spectroscopy combined with machine learning for the detection of ovarian, bowel, and peritoneal endometriosis. The Boruta algorithm was applied to identify the most informative spectral intervals for each endometriosis type, revealing characteristic wave number ranges associated with molecular changes in endometriotic tissue. For ovarian endometriosis, key intervals included 741–748 cm−1, 984–993 cm−1, and 1125–1132 cm−1 bowel endometriosis, 1055–1063 cm−1, 1077–1079 cm−1, 1561–1572 cm−1, and 1717–1720 cm−1, and for peritoneal endometriosis, 917–919 cm−1, 1542–1547 cm−1, and 1573–1576 cm−1. Three machine learning algorithms, Deep Learning (DL), Support Vector Machine (SVM), and XGBoost, were tested using both the full spectral range and the Boruta-selected feature subsets. Across all endometriosis types, XGBoost consistently outperformed DL and SVM. Using the full spectrum, XGBoost achieved accuracies of 0.81, 0.77, and 0.78 for ovarian, bowel, and peritoneal endometriosis, respectively. Feature selection with Boruta significantly improved performance, increasing accuracies to 0.93, 0.88, and 0.90, respectively, and enhancing sensitivity, specificity, precision, F1 score, MCC, and ROC AUC across all datasets. DL models often exhibited high sensitivity but poor specificity, while SVM performance improved moderately with feature selection. Overall, these results demonstrate that targeted spectral feature selection enhances the diagnostic accuracy of machine learning models for endometriosis. XGBoost, in combination with Boruta-selected spectral intervals, provides the most reliable and balanced predictions, highlighting its potential for non-invasive detection and differentiation of ovarian, bowel, and peritoneal endometriotic lesions.