MixPatch: A New Method for Training Histopathology Image Classifiers

Area of Science:

• Computational pathology
• Medical image analysis
• Machine learning in healthcare

Background:

• Convolutional Neural Networks (CNNs) are used for automated histopathological analysis.
• CNN classifiers often exhibit overconfidence, a significant issue in medical diagnostics.

Purpose of the Study:

• To introduce MixPatch, a novel training method for CNNs.
• To address prediction uncertainty and improve diagnostic performance in histopathology.

Main Methods:

• MixPatch generates a sub-training dataset with mixed-patches and soft labels per mini-batch.
• Mixed-patches are created from clean patches confirmed by pathologists.
• Proportion-based soft labeling is used for ground-truth assignment.

Main Results:

• MixPatch demonstrated superior performance and reduced overconfidence compared to other methods.
• Achieved 97.06% accuracy (1.6%–12.18% increase) and 0.76% expected calibration error (0.6%–6.3% decrease).
• Effectively handles mixed-region variations characteristic of histopathology images.

Conclusions:

• MixPatch offers a systematic approach to mitigate CNN overconfidence in medical image analysis.
• The method enhances prediction accuracy, leading to more calibrated and reliable histopathological diagnoses.