Multi-resolution deep learning characterizes tertiary lymphoid structures and their prognostic relevance in solid...

Area of Science:

• Computational pathology
• Artificial intelligence in oncology
• Immunohistochemistry and biomarker development

Background:

• Tertiary lymphoid structures (TLSs) are immune cell aggregates in tumors linked to better patient outcomes and immunotherapy response.
• Accurate TLS quantification is crucial for their use as predictive and prognostic biomarkers in solid tumors.
• Current histological assessment methods for TLS density lack standardization, hindering reliable clinical application.

Purpose of the Study:

• To develop an automated, unbiased deep learning approach for quantifying TLS and germinal centers in digital pathology slides.
• To validate the performance of the developed model against manual assessment and assess its prognostic value.

Main Methods:

• Introduction of HookNet-TLS, a multi-resolution deep learning model for automated TLS quantification.
• Development and training of HookNet-TLS using 1019 manually annotated TCGA digital pathology slides.
• Utilized datasets included clear cell renal cell carcinoma, muscle-invasive bladder cancer, and lung squamous cell carcinoma.

Main Results:

• HookNet-TLS achieved human-level performance in automated TLS quantification across multiple cancer types.
• The automated quantification demonstrated prognostic associations comparable to traditional visual assessment methods.
• The model provides objective and reproducible TLS density measurements from routine H&E stained slides.

Conclusions:

• HookNet-TLS offers a robust tool for objective TLS quantification in digital pathology, facilitating biomarker research.
• The public availability of HookNet-TLS aims to promote its widespread adoption in cancer research.
• This automated approach has the potential to standardize TLS assessment, improving biomarker reliability in clinical settings.