Classification of Malignant Lymphoma Subtypes via IHC Staining Patterns

Overview

This study proposes a novel method to improve malignant lymphoma subtype classification by evaluating the typicality of H&E-stained tissue slides based on immunohistochemical (IHC) staining patterns. Using instance selection guided by typicality measures derived from IHC stain sets, the authors demonstrate enhanced generalization ability of a multiple-instance-learning convolutional neural network (MIL-based CNN) in classifying three lymphoma subtypes.

Background

Malignant lymphomas encompass over 70 subtypes, requiring precise pathological diagnosis to guide treatment. Diagnosis typically involves initial examination of hematoxylin-and-eosin (H&E)-stained tissue slides to infer candidate subtypes, followed by immunohistochemical (IHC) staining to confirm the subtype. Digital pathology and machine learning have accelerated image analysis, but subtype classification remains challenging due to limited training data and morphological variability. This study addresses these challenges by leveraging the relationship between IHC stain usage and morphological typicality to select optimal training instances for subtype classification.

Data Highlights

Total cases analyzed: 262 malignant lymphoma cases with WSIs of H&E-stained tissue sections.

Key Findings

• The typicality of H&E-stained tissue slides can be quantitatively evaluated based on the sets of IHC stains used for definitive diagnosis.
• Instances with typical morphological features require only the common IHC stain set for diagnosis, whereas atypical cases need additional IHC stains.
• Instance selection using typicality measures improves the generalization ability of MIL-based CNN subtype classifiers.
• The study demonstrated effective three-class classification among DLBCL, AITL, and CHL using 262 clinical cases.
• Similar IHC stain sets were observed across different subtypes, and variability existed within subtypes, highlighting diagnostic complexity.

Clinical Implications

Incorporating typicality-based instance selection into digital pathology workflows can enhance machine learning model performance for lymphoma subtype classification, potentially providing pathologists with more accurate computer-aided diagnostic support. Understanding the relationship between IHC stain patterns and morphological typicality may also guide more efficient and targeted use of IHC stains in clinical practice.

Conclusion

The proposed method of evaluating H&E slide typicality via IHC stain patterns effectively improves training data selection and enhances subtype classification performance in malignant lymphoma. This approach offers a promising avenue for advancing digital pathology diagnostics through machine learning.

References

1. Author/Source/Year -- Classification of Malignant Lymphoma Subtypes Through Immunohistochemical Staining Patterns