Learnable prototype-guided multiple instance learning for detecting tertiary lymphoid structures in multi-cancer...

Area of Science:

• Pathology
• Computational Biology
• Medical Imaging

Background:

• Tertiary lymphoid structures (TLS) are critical in tumor microenvironments (TME), influencing patient prognosis and immunotherapy response.
• Accurate TLS detection in whole-slide pathological images (WSIs) is vital for clinical decisions.
• Existing multiple instance learning (MIL) methods have limitations in detecting sparse and heterogeneous TLS.

Purpose of the Study:

• To develop a weakly supervised framework for robust TLS detection in WSIs.
• To address the challenges of TLS sparsity and heterogeneity in diverse cancer types.
• To improve the generalizability of MIL for TLS analysis across different malignancies.

Main Methods:

• Proposed Learnable Prototype-Guided Multiple Instance Learning (LPGMIL) framework.
• Utilized lymphocyte-dense instances to create learnable global prototypes for feature refinement.
• Employed multiple learnable global prototypes to capture diverse TLS patterns within WSIs.
• Validated the framework on a comprehensive six-cancer-type TCGA dataset.

Main Results:

• LPGMIL demonstrated superior performance compared to existing methods on a multi-cancer dataset.
• Achieved high accuracy (76.6%), recall (74.1%), F1-score (82.7%), and AUC (83.5%).
• Effectively handled the sparsity and heterogeneity of TLS in WSIs.

Conclusions:

• LPGMIL offers an effective solution for weakly supervised TLS detection in complex cancer datasets.
• The framework enhances the analysis of TLS, crucial for predicting patient outcomes and treatment efficacy.
• This approach advances computational pathology for precision oncology.