FP-CNN: Fuzzy pooling-based convolutional neural network for lung ultrasound image classification with explainable AI

Area of Science:

• Artificial Intelligence in Medical Imaging
• Computer-Aided Diagnosis
• Medical Image Analysis

Background:

• The COVID-19 pandemic highlighted the need for efficient diagnostic tools.
• Non-invasive ultrasound imaging shows promise as a biomarker for respiratory conditions.
• Accurate classification of lung ultrasound images is crucial for timely medical diagnosis.

Purpose of the Study:

• To develop an intelligent methodology for classifying lung ultrasound images.
• To utilize a fuzzy pooling-based convolutional neural network (FP-CNN) for improved feature representation.
• To enhance diagnostic decision transparency using explainable AI (SHAP).

Main Methods:

• Development of a fuzzy pooling-based convolutional neural network (FP-CNN) for image classification.
• Implementation of Shapley Additive Explanation (SHAP) for model interpretability.
• Evaluation of various CNN architectures and fuzzy pooling strategies, including fine-tuning and multi-layer fuzzy pooling.

Main Results:

• The FP-CNN model achieved classification of ultrasound images into COVID-19, normal, and pneumonia categories.
• The Xception model, incorporating fuzzy pooling in all layers, demonstrated the highest accuracy at 97.2%.
• SHAP analysis provided explanations for the FP-CNN model's diagnostic predictions.

Conclusions:

• The proposed FP-CNN methodology offers a robust approach for diagnosing COVID-19 from lung ultrasound images.
• The integration of fuzzy pooling enhances feature extraction for better classification accuracy.
• Explainable AI methods like SHAP are vital for ensuring the trustworthiness of AI diagnostic systems.