Label-free ovarian cancer histopathological diagnosis using two-photon autofluorescence microscopy with a joint denoising and segmentation framework

Abstract

Two-photon autofluorescence (TPAF) microscopy is a promising modality for rapid, label-free assessment of unstained tissue, but is fundamentally limited by the poor visibility of nuclei, which are one of the gold-standard cancer biomarkers, due to their negative contrast nature. A restoration method capable of distinguishing these signal voids from stochastic noise to render quantitative pathology is lacking. Here, we proposed a deep learning framework of AFN-DeSeg (Auto-Fluorescence Nuclei Denoising and Segmentation) to unify image denoising and nuclear segmentation via a dual-encoder architecture incorporating a DINOv3 vision transformer. By explicitly modeling detector/shot noise while optimizing for morphological fidelity, AFN-DeSeg recovered diagnostic-grade nuclear features from noisy signals. We demonstrated that AFN-DeSeg achieved a superior reconstruction fidelity compared to state-of-the-art sequential and joint denoising-segmentation pipelines and yielded a high correlation with hematoxylin and eosin (H&E) pathology in identifying key diagnostic areas of high-grade serous ovarian carcinoma. The proposed AFN-DeSeg bridges the gap between label-free TPAF imaging and H&E-based histology, establishing a computational foundation for label-free morphological pathology and toward future intraoperative optical biopsy.