Generalizable Inter-Institutional Classification of Abnormal Chest Radiographs Using Efficient Convolutional Neural...

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Healthcare
• Computer-Aided Diagnosis

Background:

• Chest radiography is a common diagnostic tool.
• Abnormality detection in chest X-rays can be challenging.
• Deep learning models offer potential for automated analysis.

Purpose of the Study:

• Evaluate efficient Convolutional Neural Networks (CNNs) for chest radiograph abnormality detection.
• Assess the generalizability of these models on independent datasets.
• Compare the performance of DenseNet and MobileNetV2 architectures.

Main Methods:

• Trained DenseNet and MobileNetV2 models on National Institutes of Health Chest-Xray14 (NIH-CXR) and Rhode Island Hospital chest radiograph (RIH-CXR) datasets.
• Classified radiographs as normal or abnormal and predicted 14 pathological findings.
• Evaluated models using Area Under the Receiver Operating Characteristic Curve (AUROC) on internal and external test sets.

Main Results:

• Both DenseNet and MobileNetV2 achieved high AUROCs for normal/abnormal classification on both datasets (e.g., 0.960 on RIH-CXR).
• MobileNetV2 demonstrated comparable performance to DenseNet in detecting 14 pathological findings.
• External validation showed a performance decrease of 3.6-5.2% compared to locally trained models.

Conclusions:

• Efficient CNNs, like MobileNetV2, are effective for abnormality detection in chest radiographs.
• Models generalize to external data, but performance variations necessitate careful consideration for multi-institutional applications.
• AI-powered tools can aid radiologists, but require validation on diverse datasets.