Screening and early detection of cervical intraepithelial neoplasia and cervicitis using a hemoglobin absorption map-derived machine learning algorithm

Early and non-invasive detection of cervical malignancy holds great clinical significance. Diffuse reflectance (DR) spectroscopy has the capability to map tissue transformation at the biochemical, morphological, and cellular levels. We have developed a non-invasive, multimodal imaging system to map changes in tissue autofluorescence using DR for the screening and early detection of cervical cancer and cervical inflammation (cervicitis). The developed multispectral imaging device consists of light-emitting diodes (LED) emitting at 375, 545, 575, and 610 nm wavelengths, along with a 5-megapixel monochrome camera for image acquisition. Camera operation and image analysis are controlled using proprietary software installed on a Windows tablet. The 375 nm LED-excited autofluorescence, and the elastically backscattered light at 545, 575, and 610 nm originating from the cervix tissue are captured by the camera and processed to assess tissue abnormalities. A machine learning (ML) algorithm based on DR image intensity ratio values was developed for tissue classification. It was observed that the R610/R545 image ratio could discriminate malignant cervical sites from normal tissues, achieving a sensitivity of 100% and specificity of 93%. In comparison, cervicitis could be discriminated from normal tissues using the R610/R575 ratio, with a sensitivity of 91.6% and specificity of 94.4%. The study demonstrates the potential of DR imaging in conjunction with ML algorithm to non-invasively screen and detect cervical intraepithelial neoplasia and cervicitis in real time. As compared to the existing practice of Pap smear and colposcopy-directed biopsy, which are subjective and require a waiting period for results, objective screening using CerviScan would help reduce patient anxiety, unnecessary biopsies, and treatment costs. With increased patient screening, the accuracy of the ML algorithm would improve. When integrated into a cloud server, the system could address the needs of multiple users in a field setting.