Deep learning-enabled realistic virtual histology with ultraviolet photoacoustic remote sensing microscopy

Area of Science:

• Medical Imaging
• Computational Pathology
• Oncologic Surgery

Background:

• Complete tumor resection is the goal of oncologic surgeries, but positive margins are common with current histology methods.
• Frozen section analysis offers intraoperative evaluation but has known inaccuracies.
• Accurate margin assessment is critical for effective cancer treatment and patient outcomes.

Purpose of the Study:

• To introduce a novel, label-free histological imaging method for intraoperative margin assessment.
• To combine ultraviolet photoacoustic remote sensing and scattering microscopy with deep learning for virtual staining.
• To evaluate the diagnostic accuracy and pathologist preference of the new method compared to conventional techniques.

Main Methods:

• Developed a label-free histological imaging system using ultraviolet photoacoustic remote sensing and scattering microscopy.
• Employed unsupervised deep learning (cycle-consistent generative adversarial network) for realistic virtual staining of unstained tissues.
• Quantitatively validated the method against H&E-stained histology in prostate and breast tissues.
• Conducted diagnostic utility studies and a blinded pathologist survey.

Main Results:

• The imaging method scans unstained tissues at up to 7 mins/cm², achieving resolution equivalent to 400x digital histopathology.
• Strong concordance was observed with conventional histology in benign and malignant prostate and breast tissues.
• Achieved high diagnostic performance: mean sensitivity of 0.96 and specificity of 0.91 for breast, and 0.87 and 0.94 for prostate specimens.
• Pathologists preferred the virtual stain quality over frozen section analysis (P=0.03).

Conclusions:

• The novel label-free imaging technique provides accurate and rapid intraoperative margin evaluation in oncologic surgery.
• This method, utilizing advanced microscopy and AI, offers a promising alternative to conventional histology and frozen sections.
• The improved virtual staining and diagnostic accuracy have the potential to enhance surgical outcomes by reducing positive margins.