A loss-based patch label denoising method for improving whole-slide image analysis using a convolutional neural...

Area of Science:

• Computational pathology
• Medical image analysis
• Deep learning applications in oncology

Background:

• Automated whole-slide image classification is crucial for cancer diagnosis but hindered by noisy labels.
• Pathologist annotations of malignant regions can inadvertently include benign areas, leading to low accuracy and Type-II errors.

Purpose of the Study:

• To develop a deep learning-based method for denoising patch labels in whole-slide images.
• To enhance the accuracy of automated cancer classification despite inherent annotation noise.

Main Methods:

• Proposed a novel deep learning-based patch label denoising method named LossDiff.
• Utilized convolutional neural networks (CNNs) for noisy patch classification.
• Validated the method on stomach cancer images and publicly available datasets at various noise levels.

Main Results:

• Achieved high classification accuracies: 98.81% (binary), 97.30% (ternary), and 89.47% (quaternary classes).
• Demonstrated significant improvement over existing methods in patch-based cancer classification.
• Confirmed the robustness of the LossDiff method under different noise conditions.

Conclusions:

• The LossDiff method offers a simple yet effective solution for noisy patch classification in whole-slide images.
• The approach has practical implications for improving whole-slide image annotation and automated cancer diagnosis.
• Addresses the challenges of high annotation costs and human error in medical image labeling.