Improved Explainability for Quality Classifiers Demonstrated in Ultrasound Medical Imaging

This study introduces a novel refinement to two classes of local explainability techniques and combines this with the Self-Organizing Map (SOM) to achieve a combination of local and global explainability. The new approach is demonstrated on a deep neural network that has been trained to classify the quality of a certain type of ultrasound exam. Using the new approach, we are able to demonstrate that the deep network uses two different types of image characteristics in combination to assess quality. This insight would not have been possible using standard local explainability methods. In a clinical setting, the novel explanations can enhance confidence in the deep network decisions.