Explainable DCNN Decision Framework for Breast Lesion Classification from Ultrasound Images Based on Cancer...

Area of Science:

• Medical Image Analysis
• Artificial Intelligence in Healthcare
• Radiology

Background:

• Deep convolutional neural networks (DCNNs) excel in medical image analysis, particularly for breast lesion classification in ultrasound (US) images.
• Explainability of DCNN decisions is crucial for healthcare system adoption and trust.
• Current DCNN solutions lack transparent decision-making processes in medical diagnostics.

Purpose of the Study:

• To present a novel framework for explaining DCNN classification decisions in ultrasound images.
• To link DCNN-derived saliency maps to established cancer characteristics for improved interpretability.
• To demonstrate the framework's utility in breast lesion classification.

Main Methods:

• Developed DCNN models for ultrasound image classification using a transfer learning approach.
• Applied visualization methods (EGrad-CAM, Ablation-CAM) to generate saliency maps.
• Mapped saliency map outputs to known cancer characteristics (echogenicity, calcification, shape, margin) for benign and malignant lesions.

Main Results:

• The developed framework successfully explained DCNN classification decisions for breast lesions.
• Saliency maps revealed distinct contributions of characteristics like echogenicity, calcification, shape, and margin to lesion classification.
• The study utilized a retrospective dataset of 1298 ultrasound images for evaluation.

Conclusions:

• The proposed framework enhances the explainability of DCNN models in medical image analysis.
• Understanding the contribution of specific characteristics improves trust and acceptance of AI in diagnostics.
• This work lays the groundwork for explaining DCNN decisions in other cancer types.