Detection and visualization of abnormality in chest radiographs using modality-specific convolutional neural network...

Area of Science:

• Medical Imaging and Artificial Intelligence
• Computer-Aided Diagnosis

Background:

• Convolutional Neural Networks (CNNs) excel at image tasks but require adaptation for unique medical image features.
• Training on large, modality-specific datasets improves generalization for medical recognition tasks.

Purpose of the Study:

• To enhance abnormality detection and localization in chest X-rays (CXRs) using modality-specific ensemble learning.
• To improve the performance and generalization of CNN models for CXR analysis.

Main Methods:

• CNN models were trained on a large CXR dataset to learn modality-specific features.
• Ensemble strategies combined predictions from multiple CNNs to reduce variance and improve robustness.
• Class-selective relevance mapping (CRM) was employed for visualizing model predictions and localizing abnormalities.

Main Results:

• Ensemble models demonstrated superior performance compared to individual CNNs in detecting and localizing abnormalities.
• Localization accuracy was significantly improved, as indicated by Intersection of Union (IoU) and mean Average Precision (mAP) metrics.
• CRM effectively visualized discriminative regions of interest, providing better model interpretability.

Conclusions:

• Modality-specific ensemble learning is a promising approach for improving abnormality detection in CXRs.
• Ensemble methods enhance CNN generalization and robustness in medical image analysis.
• CRM aids in understanding and validating CNN predictions for clinical applications.