Computer-Aided System for the Detection of Multicategory Pulmonary Tuberculosis in Radiographs.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pulmonology

Background:

• Early screening and diagnosis of tuberculosis (TB) are crucial for effective control and treatment.
• Current methods for TB detection in chest radiographs have limitations in specificity and sensitivity for various lesion types.

Purpose of the Study:

• To develop an integrated computer-aided system (CAS) utilizing deep learning for the detection of multiple categories of tuberculosis lesions in chest radiographs.
• To enhance the accuracy and efficiency of TB screening and diagnosis, particularly in endemic regions.

Main Methods:

• A fully convolutional neural network (FCNN) was employed to segment lung areas for pulmonary TB detection.
• A novel multicategory tuberculosis lesion detection method was proposed, incorporating a learning scalable pyramid structure into the Faster Region-based Convolutional Network (Faster RCNN).
• Reinforcement learning was utilized to minimize false-positive detections.

Main Results:

• The proposed Faster RCNN with a scalable pyramid structure improved the detection of small-area lesions and identified challenging samples during training.
• The system achieved high performance on public datasets: Montgomery (AUC = 0.977, accuracy = 0.926) and Shenzhen (AUC = 0.941, accuracy = 0.902).
• A classification rule for whole chest X-rays was developed, demonstrating the model's effectiveness.

Conclusions:

• The developed deep learning-based CAS demonstrates superior performance compared to existing systems for tuberculosis detection.
• The system effectively assists radiologists in diagnosis and aids public health providers in TB screening, especially in endemic areas.