US20250278839
2025-09-04
Physics
G06T7/11
The disclosed method and system utilize deep learning to digitally stain microscopic images from label-free samples. This approach is particularly useful for creating whole slide images (WSIs) of unstained tissue samples, which are then analyzed by histopathologists. By bypassing traditional histochemical staining processes, this method saves time and reduces costs. A convolutional neural network, trained with a generative adversarial network model, transforms fluorescence images of unlabeled samples into images equivalent to those obtained from chemically stained samples.
The invention relates to imaging methods for unstained tissue using deep neural network learning. It focuses on microscopy systems that apply deep learning algorithms to digitally stain images of label-free tissue sections. The technique uses endogenous fluorescence signals to map functional and structural properties of biological specimens, providing a digital alternative to traditional staining methods.
Traditional tissue imaging involves a labor-intensive process involving formalin fixation, paraffin embedding, sectioning, staining, and brightfield microscopy. This process uses multiple reagents and affects the tissue irreversibly. Recent advances have explored different imaging modalities such as non-linear microscopy and fluorescence techniques. However, these often require complex equipment like ultra-fast lasers and are not widely accessible. The fluorescence signal offers unique imaging opportunities by utilizing light emitted from endogenous fluorophores in the tissue.
In one embodiment, a deep neural network is used for digital staining of label-free samples using their fluorescence images. This network is trained with matched pairs of chemically stained images and corresponding fluorescence images. By inputting a fluorescence image of an unstained sample into the network, a digitally stained image is generated quickly, replacing traditional staining and imaging steps. The method has been validated on various human tissues, producing results comparable to conventional stains like H&E, Jones stain, and Masson's Trichrome.
The digital staining method can be applied across different excitation wavelengths and microscopy modalities. It allows for rapid diagnosis in clinical settings such as surgery rooms or biopsy sites and supports telepathology applications. Beyond clinical use, it benefits histology research and education by preserving unlabeled tissue sections for further analysis. Pathologists have validated the technique's efficacy through blind evaluations, recognizing histopathological features with high agreement to traditionally stained images.