Explaining a Deep Learning Based Breast Ultrasound Image Classifier with Saliency Maps.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Deep neural networks (DNNs) demonstrate high performance in breast mass classification from ultrasound (US) images.
• Clinical adoption of DNNs is hindered by their 'black box' nature, lacking decision explainability.

Purpose of the Study:

• To address the explainability challenge in DNN-based breast mass classification.
• To generate saliency maps that highlight critical regions in US images influencing classification decisions.

Main Methods:

• A DNN model was developed using transfer learning on 272 breast mass US images (123 malignant, 149 benign).
• Class activation mapping (CAM) generated saliency maps for each image.
• The pointing game metric quantitatively assessed saliency map overlap with predefined image regions (mass, peritumoral, below mass).

Main Results:

• The DNN classifier achieved an area under the ROC curve of 0.887, accuracy of 0.835, sensitivity of 0.801, and specificity of 0.868.
• Saliency map analysis revealed that network decisions correlated with specific tissue regions in 71% of correctly classified cases.

Conclusions:

• This study enhances the understanding of DNN models for breast mass diagnosis.
• Demonstrated that DNN classification decisions are associated with specific tissue characteristics visible in US images.