Second-order multi-instance learning model for whole slide image classification.

Area of Science:

• Computational pathology
• Digital pathology
• Machine learning in oncology

Background:

• Whole slide histopathology images (WSIs) are vital for diagnosing breast cancer lymph node metastasis.
• WSIs often lack detailed tumor annotations and possess very large resolutions.
• Multi-instance learning (MIL) is a key framework for WSI classification with only slide-level labels.

Purpose of the Study:

• To develop a novel second-order multiple instance learning (SoMIL) method for histopathological image classification.
• To enhance the analysis of large-resolution WSIs lacking fine-grade annotations.
• To improve the accuracy of diagnosing lymph node metastasis in breast cancer.

Main Methods:

• Developed a second-order multiple instance learning (SoMIL) framework.
• Employed a second-order pooling module (matrix power normalization covariance) for instance-level feature extraction.
• Integrated an attention mechanism and a recurrent neural network (RNN) for adaptive feature aggregation.

Main Results:

• SoMIL demonstrated significant improvement over state-of-the-art MIL methods on the 2016 Camelyon grand challenge dataset.
• Achieved competitive area under the curve performance compared to fully-supervised methods in external validation.
• Effectively explored second-order statistics of deep features for improved classification.

Conclusions:

• The proposed SoMIL framework offers a powerful approach for weakly supervised histopathological image classification.
• SoMIL enhances diagnostic accuracy for breast cancer lymph node metastasis.
• This method shows promise for analyzing large-scale WSIs in digital pathology.