Abstract
Cervical cancer originates from precursor lesions in the cervix, where early diagnosis is essential to avoid progression to invasive carcinoma. Today, lesion detection relies mainly on colposcopy, a highly operator-dependent procedure with considerable variability in accuracy, particularly among less-experienced clinicians. In this study, we present a hyperspectral imaging-based approach for assisted colposcopy. We developed a custom hyperspectral colposcope covering 470-900 nm and conducted a 32-month clinical study including 116 patients and 245 hyperspectral images acquired during routine examinations. First, a dedicated algorithm automatically delineated the cervical region and segmented the tissue into ectocervix, endocervix and abnormal areas, achieving a macro-level DICE of 0.84. Subsequently, two approaches were studied for pixel-wise lesion classification: a binary model for high-grade squamous intraepithelial lesion and invasive carcinoma versus healthy tissue, which achieved a mean F1-score of 0.74 on an independent test set; and a multiclass model for grading according to the Bethesda system, which showed lower generalisation (F1-score = 0.26) due to limited spectral resolution and spectral overlap. Overall, the results show the potential of hyperspectral colposcopy for non-invasive detection and delimitation of cervical lesions.