Subtype classification of malignant lymphoma using immunohistochemical staining pattern.

Area of Science:

• Computational pathology
• Digital pathology
• Machine learning in medicine

Background:

• Accurate image classification in pathology is crucial for diagnosis.
• Limited training data can hinder classifier generalization ability.
• Hematoxylin and eosin (H&E) staining is standard, but immunohistochemical (IHC) stains provide specific molecular information.

Purpose of the Study:

• To propose a method for quantitatively evaluating the typicality of H&E-stained tissue slides based on IHC staining patterns.
• To apply this typicality measure for instance selection in training classifiers for malignant lymphoma subtype prediction.
• To improve the generalization ability of classifiers, particularly when training data is scarce.

Main Methods:

• Defined tissue slide typicality using the probability density ratio of IHC staining patterns in a low-dimensional embedded space.
• Employed a multiple-instance-learning-based convolutional neural network for subtype classification without region annotations.
• Selected training instances based on the computed typicality to enhance classifier generalization.

Main Results:

• Typical instances showed more accurate subtype prediction compared to atypical instances.
• Instance selection based on typicality improved classifier generalization ability.
• Classification accuracy increased from 0.664 (baseline) to 0.683 when focusing on typical instances for training.

Conclusions:

• Typicality of H&E slides, derived from IHC patterns, serves as a valuable criterion for instance selection.
• This approach effectively enhances classifier generalization ability in malignant lymphoma subtype classification.
• The proposed typicality measure is practical for instance selection, even with limitations.