A Polish research team has shown that different forms of endometriosis may be distinguished using tissue biochemistry measured by FTIR spectroscopy combined with machine learning, potentially reducing reliance on invasive diagnostic procedures.
In a study published in Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy, the researchers found that biochemical “fingerprints” of tissue samples, analysed with Fourier-transform infrared (FTIR) spectroscopy and advanced algorithms, allowed them to differentiate between types of endometriosis with up to 93% accuracy after feature selection. Endometriosis is a chronic inflammatory disease that often remains undiagnosed for years and significantly affects patients’ quality of life.
The study authors note that, on average, 7–10 years pass between the onset of symptoms and diagnosis, and that about 50% of patients experience difficulties conceiving. Endometriosis occurs when tissue similar to the uterine lining develops outside the uterus, where it can grow, bleed cyclically and trigger inflammation and fibrosis, leading to adhesions, chronic pain and impaired organ function.
The disease can take several forms, including superficial peritoneal lesions, ovarian cysts and deeply infiltrating endometriosis, in which tissue penetrates more than 5 mm. The latter is considered the most aggressive and distressing form.
Scientists from Lublin, Rzeszów and Kraków set out to determine whether these forms could be identified not only through laparoscopy and histopathology, but also through biochemical analysis. FTIR spectroscopy measures how tissue absorbs infrared light, producing a spectrum that reflects its molecular composition. According to the authors, this method can reveal chemical differences between healthy and diseased tissue that are not always visible under a microscope.
Tissue samples were collected during laparoscopic surgeries at the endometriosis treatment centre of the 1st Military Clinical Hospital in Lublin. For each patient, samples from endometriotic lesions and from healthy-looking tissue were taken as controls and prepared as thin sections for FTIR analysis.
The spectra showed repeatable differences between healthy and diseased tissue, including stronger lipid-related signals and changes in protein bands, as well as alterations in regions associated with sugars, nucleic acids and phosphate compounds. However, because each spectrum contains thousands of data points, the researchers applied machine learning to identify diagnostic patterns.
They compared three approaches: Support Vector Machines (SVM), Deep Learning (DL) and XGBoost. While DL models achieved perfect sensitivity for some lesion types, they performed poorly in recognising negative cases. SVM proved robust but less effective with noisy, high-dimensional data. XGBoost, which combines multiple decision trees, delivered the most balanced and reliable results.
To further improve performance, the team used the Boruta algorithm to select only the most informative spectral features. With full spectra, XGBoost achieved accuracies of about 81% for ovarian endometriosis, 77% for intestinal endometriosis and 78% for peritoneal endometriosis. After feature selection, accuracy rose to 93%, 88% and 90%, respectively, with improved balance between sensitivity and specificity.
The analysis also identified different key spectral regions for lesions in the ovary, intestine and peritoneum, supporting the conclusion that these forms of endometriosis are biochemically distinct.
The authors argue that combining FTIR spectroscopy with carefully selected algorithms could provide rapid, objective diagnostic support, help differentiate lesion types for treatment planning, shorten the diagnostic process and, in the future, reduce the need for invasive laparoscopy. (PAP)
PAP - Science in Poland
kmp/ agt/
tr. RL
