Automating the Detection of Acetowhite Lesions by Classifying the Temporal Behavior of Cervical Regions

Cervical acetowhitening is a strong visual cue for lesion identification; however, clinician opinions are subjective and inconsistent. To provide objective evaluation, a machine learning model that enhances colposcopy by automating acetowhite lesion detection in cervical time-series images was developed.

A dataset of time-series images collected during 238 colposcopy examinations, where in each case, acetowhitening areas had been annotated by 5 expert colposcopists, was utilized. After preprocessing the images, which included aligning them and extracting the cervix by bounding-box detection, the time-series images were divided into discrete segments, and the transient visual changes were extracted as local statistical features. Using the local agreement levels of the reviewers' predictions for the presence of a lesion, a machine learning model was trained to classify the time-series data and generate synthetic annotations.

The model-generated synthetic acetowhite lesion maps aligned well with the experts' annotations. In a subset of cases with localized histology results from punch biopsy, model predictions correlated well with histopathology, highlighting potential for clinical utility.

The proposed method was effective in mimicking expert colposcopists in identifying lesions based on classification of time-series images and outputs correlated with histology. Application of a machine learning model in the clinical setting may streamline lesion identification, support clinician decision-making for punch biopsy, and eventually improve accuracy in detecting precancerous lesions.