Sensitivity analysis for interpretation of machine learning based segmentation models in cardiac MRI.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Artificial neural networks automate medical image segmentation, but their performance degrades with data variations.
• Neural networks are often black boxes, making it difficult to understand their decision-making processes and predict generalization.
• Interpreting segmentation models is crucial for reliable application in clinical settings.

Purpose of the Study:

• To present a generic method for interpreting segmentation models using sensitivity analysis.
• To introduce an open-source Python library (misas) for implementing sensitivity analysis.
• To demonstrate the utility of sensitivity analysis for evaluating model suitability and robustness.

Main Methods:

• Sensitivity analysis was employed, involving controlled modification of model inputs to evaluate effects on segmentation outputs.
• An open-source Python library, misas, was developed to facilitate sensitivity analysis for various models and datasets.
• Two case studies using cardiac magnetic resonance imaging data were conducted to showcase the method's application.

Main Results:

• Sensitivity analysis provides insights into model sensitivity to input alterations and feature importance.
• The misas library enables practical application of sensitivity analysis for segmentation models.
• Case studies demonstrated the suitability of a model for a new dataset and evaluated the robustness of a newly trained model.

Conclusions:

• Sensitivity analysis enhances the interpretability of segmentation models for developers and clinicians.
• Understanding neural network behavior through sensitivity analysis aids in decision-making regarding model deployment.
• The presented approach and software are broadly applicable beyond cardiac MRI segmentation.