Human-Machine Integration to Enhance Clinical Typing and Fine-Grained Interpretation of Cervical OCT Images

Cervical cancer is the fourth most common cancer among women worldwide, posing a severe threat to female reproductive health. The cervical optical coherence tomography (OCT) technology, known for its high-resolution imaging and non-invasive characteristics, holds significant potential in gynecology applications. Annotating cervical OCT images is based on pathological typing of cervical tissue from matching OCT images with the corresponding pathological slice images. It depends mostly on domain-specific skills and experience, lacking comprehensive, straightforward interpretation of subtypes for cervical OCT images. The accuracy of computer-aided diagnosis for cervical OCT images is always limited, failing to meet gynecologists’ clinical requirements. Therefore, our work aims to establish a fine-grained subtyping method for cervical OCT images to complement high-level pathological categories. The proposed method is implemented in a self-supervised pre-training style to explore OCT image characteristics fully. Unlike data-driven clustering methods, our method constructs a form of anchor points to leverage medical prior knowledge to avoid the mismatch between the discovered rule and existing pathological knowledge. We assessed our method with competitive baseline approaches on internal and external datasets. Besides, we collaborated with medical experts to demonstrate that the image feature patterns discovered are normative and clinically valuable for gynecologists.