Researchers have developed AFN-DeSeg, a deep learning framework that correlates directly with hematoxylin and eosin pathology in identifying key diagnostic areas of high-grade serous ovarian carcinoma. The work addresses a limitation of two-photon autofluorescence microscopy, a promising method for rapid, label-free tissue assessment, by improving the visibility of nuclei, a cancer biomarker historically difficult to detect due to naturally low contrast. By unifying image denoising and nuclear segmentation, AFN-DeSeg recovers diagnostic-grade nuclear features and “bridges the gap between label-free TPAF imaging and H&E-based histology,” establishing a computational foundation for potential intraoperative optical biopsy.
Two-Photon Autofluorescence Microscopy for Ovarian Cancer Histopathology
Standard segmentation approaches for identifying cancerous cells often depend on clear boundaries, yet two-photon autofluorescence (TPAF) imaging presents challenges due to the negative contrast of nuclei and a low signal-to-noise ratio that obscures their morphology. To overcome these limitations, researchers constructed a composite dataset using both public fluorescence microscopy images and in-house ovarian cancer TPAF images, addressing the scarcity of paired TPAF data needed for training advanced analytical frameworks.
This strategy allowed training on realistic detector-noise statistics specific to their two-photon imaging system while incorporating the morphological diversity of established biological datasets, as demonstrated by images shown in Fig. S8a. Deep learning efficacy in microscopy is frequently limited by the lack of paired “noisy–clean” reference data, a problem the team directly addressed through this synthetic data generation. This was a key component of validating the framework’s ability to accurately identify cellular structures within the complex TPAF imagery.
Beyond image enhancement, the framework’s performance was rigorously tested against established diagnostic standards. This capability could enable faster diagnoses during surgery, bypassing the delays associated with traditional pathology methods. Further validation included confirming the framework’s ability to improve image clarity and facilitate accurate analysis. The team’s work builds upon earlier advances in microscopy, establishing a computational foundation for label-free morphological pathology, and toward future intraoperative optical biopsy.
AFN-DeSeg: Joint Denoising and Segmentation Framework for TPAF Images
The AFN-DeSeg framework uses the DINOv3 vision transformer, integrating it within a dual-encoder architecture to simultaneously address image denoising and nuclear segmentation, a combined approach the researchers claim surpasses existing sequential or joint pipelines. This architecture incorporates an instrument-calibrated Mixed Poisson-Gaussian (MPG) noise model, allowing the network to learn the statistical properties of two-photon autofluorescence (TPAF) imaging and bridge the gap between synthetic and real-world data.
The framework’s design prioritizes efficient capture of the unique texture of ovarian tissue, a critical element in accurate diagnosis. To establish a robust validation process, the team created a benchmarking foundation using a human-in-the-loop annotation workflow, ensuring high-quality ground truth data for comparison. AFN-DeSeg accepts noisy TPAF images as input and generates instance segmentation masks, delineating individual nuclei within the tissue sample.
Performance was assessed against established methods, including sequential denoising-segmentation and other joint approaches like DenoiSeg and Cellpose 3, specifically on TPAF images of ovarian cancer tissue where accurate nuclei identification is inherently challenging due to signal voids, irregular shapes, and the MPG noise. This formed a key part of this validation, according to the published work.
The framework’s efficacy extends beyond image enhancement; the resulting nuclear masks were directly used to predict key pathological diagnostic areas, achieving high concordance with hematoxylin and eosin (H&E) histopathology standards. The researchers utilized both public image datasets and an in-house ovarian cancer cohort for training and evaluation, ensuring a comprehensive assessment of the framework’s capabilities.
The team also utilized a separate cohort to further refine the framework’s predictive capabilities. These were employed to confirm the framework’s reliability in a clinical setting, demonstrating its potential to aid pathologists in making more informed and timely diagnoses.
DINOv3 Vision Transformer Captures Ovarian Tissue Texture
The AFN-DeSeg framework uses a DINOv3 vision transformer to capture nuanced ovarian tissue textures, a capability important for accurate analysis when using label-free two-photon autofluorescence microscopy. Unlike conventional convolutional networks that focus on local edge detection, the DINOv3 stream employs self-attention mechanisms to discern global features, enabling the model to infer the presence of nuclei even in high-noise regions where local gradients are unreliable. This approach addresses a key limitation of two-photon imaging, historically challenged by visualizing nuclei, due to their naturally low contrast against surrounding tissue.
The integration of this transformer with a U-Net encoder, facilitated by a LoRA Adapter & Fusion module, allows for efficient adaptation to the specific characteristics of TPAF images. AFN-DeSeg’s dual-encoder architecture simultaneously recovers high signal-to-noise ratio TPAF images and generates precise nuclear masks, a process demonstrated to outperform existing sequential and joint denoising-segmentation pipelines.
The framework’s ability to accurately delineate varied nuclear shapes and sizes, and to separate aggregated nuclei instances, substantially reduces false-positive segmentation of non-nuclei vesicles. This simultaneous recovery of image quality and nuclear detail is attributed to the DINOv3 transformer’s capacity to infer structural continuity based on contextual information, even when pixel data is corrupted. “Specifically, AFN-DeSeg accurately delineates the varied shapes and sizes of tumor nuclei atypia,” the researchers state, further emphasizing the framework’s precision in identifying subtle but significant morphological features.
AFN-DeSeg Correlates with Hematoxylin and Eosin Pathology
Evaluation of AFN-DeSeg nuclei segmentation performance on two-photon autofluorescence (TPAF) images of ovarian cancer tissue revealed a high degree of consistency with hematoxylin and eosin (H&E) histology and established cancer pathology features, demonstrating a potential for improved diagnostic accuracy. The framework addresses a persistent challenge in TPAF imaging, the difficulty in visualizing nuclei due to their naturally low contrast, by simultaneously recovering both image quality and detailed nuclear morphology.
Baseline methods often misinterpret background noise as nuclear voids, leading to over-segmentation, but AFN-DeSeg mitigates this issue through explicit modeling of detector and shot noise while prioritizing morphological fidelity. Comparative analysis against the Cellpose benchmark and H&E references showed AFN-DeSeg’s ability to handle complex morphologies more effectively, while also yielding a high correlation with standard pathology assessments.
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. Pathologist-corrected attention heatmaps and CLAM-generated visualizations further highlight the diagnostic concordance between AFN-DeSeg’s TPAF predictions and H&E-based assessments; green areas indicate overlap, while blue and red denote modality-specific discrepancies.
The researchers note that the current framework was tested on FFPE-sectioned tissue, the standard preparation for computational pathology, allowing for direct spatial correlation with H&E results; unfixed fresh tissue presents different challenges due to variations in autofluorescence intensity, tissue thickness, and nuclear morphology.
Despite these limitations, the work establishes a computational foundation for translational studies and opens the door to simultaneous acquisition of TPAF intensity for NADH and FAD, potentially enabling in vivo dual-modal diagnosis combining morphological pathology with optical redox ratio (ORR) imaging. “The proposed AFN-DeSeg establishes the computational foundation for the translational study,” the authors state, suggesting a pathway toward and faster diagnoses during surgery.
Addressing Limitations of Sequential Denoising-Segmentation Pipelines
Traditional sequential denoising-segmentation pipelines, where noise reduction precedes nuclear identification, often propagate errors by mistaking cytoplasmic noise for nuclear voids, leading to over-segmentation, a problem AFN-DeSeg circumvents through its unified approach. This is particularly important given that accurate nuclei delineation is inherently difficult in TPAF imaging due to the negative contrast and irregular shapes common in cancerous tissues. The architecture of AFN-DeSeg directly addresses these limitations through the integration of a DINOv3 vision transformer, a component that provides a global semantic context lacking in conventional convolutional neural networks.
While U-Net encoders excel at local edge detection, they struggle to differentiate between signal-dependent speckle noise and genuine tissue texture in low-contrast areas; the transformer backbone, pre-trained with visual priors, infers structural continuity even with corrupted pixel information. The system’s capacity to simultaneously acquire TPAF intensity of NADH and FAD opens possibilities for obtaining metabolic images via optical redox ratio (ORR), enabling the concurrent assessment of both structural and metabolic information.
Despite its advancements, the framework acknowledges certain limitations, notably its reliance on a specific Poisson-Gaussian noise model calibrated to the researchers’ particular photomultiplier tube (PMT) detectors. However, the team anticipates that future iterations could incorporate more adaptable noise models, broadening the framework’s applicability across diverse imaging setups and tissue types.
Potential for Intraoperative Optical Biopsy with Label-Free TPAF
The AFN-DeSeg framework yielded a high correlation with hematoxylin and eosin (H&E) pathology in identifying key diagnostic areas of high-grade serous ovarian carcinoma. Current diagnostic methods, like Hematoxylin and Eosin (H&E) histopathology, are ex vivo and time-consuming, unsuitable for real-time surgical decision-making; surgeons often rely on frozen section analysis, a technique prone to artifacts and sampling errors. This new system aims to provide rapid, label-free histopathological assessment directly within the operating room, addressing an urgent clinical need for more reliable biopsy technologies.
Researchers found variance increases slightly in dense tumor regions due to nuclei aggregation, but the model maintained robustness without saturation, demonstrating its adaptability to complex tissue structures. Beyond accuracy, AFN-DeSeg’s design focuses on overcoming the trade-off between acquisition speed and image quality inherent in TPAF.
High-speed acquisition, necessary for intraoperative use, inevitably reduces photon density and degrades the visibility of cell nuclei, a key diagnostic feature. “Our findings suggest that one major barrier to real-time optical biopsy is computational,” the study reports, “and the gap between label-free optical signals and H&E-based pathology can be efficiently bridged by integrating foundation models with domain-specific restoration tasks.” This integration, utilizing a DINOv3-based encoder, allows for reconstruction of diagnostic-grade nuclear morphology even from photon-starved, noisy acquisitions of formalin-fixed and paraffin-embedded (FFPE) ovarian cancer tissue.
The team acknowledges that future work should evaluate the transferability of the DINOv3 encoder to other tissue types, determining if the visual priors learned represent universal histological features. This computational foundation, coupled with the framework’s demonstrated accuracy, positions AFN-DeSeg as a significant advancement toward realizing the potential of real-time, label-free optical biopsy for improved cancer diagnosis and surgical outcomes.
See today’s quantum computing news on Quantum Zeitgeist for the latest breakthroughs in qubits, hardware, algorithms, and industry deals.
