Lung Cancer Diagnosis on Virtual Histologically Stained Tissue Using Weakly Supervised Learning

Area of Science:

• Computational pathology
• Artificial intelligence in oncology
• Digital pathology and medical imaging

Background:

• Lung adenocarcinoma (LUAD) is the most common lung cancer, typically diagnosed via manual interpretation of H&E-stained tissue slices.
• Current histological staining and diagnostic procedures are labor-intensive, time-consuming, and require repetitive manual interpretation.
• Deep learning offers potential for automated cancer detection on histology images, but traditional workflows remain a bottleneck.

Purpose of the Study:

• To develop a weakly supervised learning method for LUAD classification using label-free tissue slices with virtual histological staining.
• To enable rapid, cost-effective, and interpretable pathological examination by converting autofluorescence images to virtual H&E-stained images.
• To assess the performance of an attention-based multiple-instance learning model on virtual H&E-stained whole-slide images (WSIs).

Main Methods:

• A weakly supervised deep generative model was employed to convert autofluorescence images of label-free tissue into virtual H&E-stained images.
• An attention-based multiple-instance learning model was trained for LUAD classification on both open-source and collected H&E-stained WSIs.
• The model's performance was validated using Area Under the Curve (AUC) metrics on standard H&E-stained and virtual H&E-stained WSIs.

Main Results:

• The model achieved high performance on standard H&E-stained WSIs (AUC of 0.961) and comparable results on virtual H&E-stained WSIs (AUC of 0.973 vs. 0.977).
• Attention heatmaps generated from virtual H&E-stained WSIs effectively indicated tumor regions, similar to ground-truth H&E images.
• The virtual staining approach demonstrated strong diagnostic capability, comparable to traditional methods.

Conclusions:

• The proposed diagnostic workflow using virtual H&E-stained images of label-free tissue provides a rapid, cost-effective, and interpretable alternative for pathological examination.
• This method can assist clinicians in postoperative pathological assessments and has the potential to be adapted for other imaging modalities and diseases.
• The study highlights the feasibility of deep learning and virtual staining for efficient and accurate LUAD classification.