DIOR-ViT: Differential ordinal learning Vision Transformer for cancer classification in pathology images

Area of Science:

• Computational pathology
• Machine learning
• Medical image analysis

Background:

• Cancer grading is crucial for treatment but often treated as simple classification.
• Existing methods ignore the inherent order (e.g., higher grade = worse prognosis).
• This limits the full utilization of information in cancer grade labels.

Purpose of the Study:

• To introduce a novel differential ordinal learning framework for cancer grading.
• To develop a transformer-based neural network for simultaneous classification and ordinal learning.
• To improve the accuracy and reliability of computational cancer grading.

Main Methods:

• Defined differential ordinal learning to quantify differences between sample class labels in feature space.
• Proposed a transformer neural network architecture for integrated classification and ordinal learning.
• Developed a specialized loss function tailored for differential ordinal learning.

Main Results:

• The proposed method demonstrated improved accuracy and reliability in cancer grading across three distinct datasets.
• Differential ordinal learning outperformed conventional cancer grading approaches.
• The approach effectively leverages the ordinal nature of cancer grades.

Conclusions:

• Differential ordinal learning offers a more robust approach to cancer grading in computational pathology.
• The transformer-based method enhances grading accuracy by incorporating label order.
• This framework has potential applications in other disease grading with ordinal relationships.