Global Explainability of A Deep Abstaining Classifier for Cancer Pathology Reports.

Area of Science:

• Computational pathology
• Artificial intelligence in medicine
• Machine learning for cancer diagnosis

Background:

• Multitask deep abstaining classifiers (DAC) are used for automated cancer pathology report annotation.
• These models predict multiple cancer attributes simultaneously but may abstain on ambiguous cases, reducing coverage.
• Achieving high accuracy (e.g., 97%) on histology prediction led to retaining only 22% of samples.

Purpose of the Study:

• To present a global explainability method for characterizing error sources in a real-world multitask DAC.
• To identify specific reasons for classification errors in cancer histology prediction.
• To suggest strategies for improving the DAC's performance and coverage.

Main Methods:

• Applied a global explainability method using dimensionality reduction on aggregated local explanations (ALE).
• Utilized Grad-CAM for local explainability on ~1.04 million annotated cancer pathology samples.
• Analyzed error sources globally, including hierarchical complexity, label noise, and insufficient information.

Main Results:

• The global explainability method successfully identified key error sources in histology classification.
• Identified error sources include hierarchical class complexity, label noise, insufficient evidence, and conflicting data.
• The method provides a tractable approach to global explainability for complex deep learning models.

Conclusions:

• The developed global explainability method offers insights into DAC errors for cancer histology.
• Strategies for iterative improvement include refining exclusion criteria and focused annotation.
• Reducing penalties for errors involving hierarchically related classes can enhance model performance.