Intelligent diagnosis model for chest X-ray images diseases based on convolutional neural network

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Multi-label medical image classification faces challenges due to feature coupling, leading to misdiagnoses.
• Accurate identification of complex lesions in chest X-rays is crucial for effective clinical decision-making.

Purpose of the Study:

• To develop a chest X-ray pathology reasoning method that addresses feature coupling in multi-label classification.
• To enhance the precise identification of complex lesions and improve clinical disease prediction accuracy.

Main Methods:

• Utilized hierarchical attention convolutional networks and a multi-label decoupling loss function.
• Incorporated adaptive dilated convolution with deformable kernels for multi-scale feature extraction.
• Employed a channel-space dual-path attention mechanism for lung field partitioning and lesion localization.
• Implemented cross-scale skip connections to fuse multi-level feature information.
• Applied a KL divergence-constrained contrastive loss function with orthogonal regularization to decouple pathological features.

Main Results:

• Achieved a weighted F1-score of 0.97 on the ChestX-ray14 dataset.
• Reported a Hamming Loss of 0.086.
• Demonstrated AUC values exceeding 0.94 for all pathologies, indicating high diagnostic performance.

Conclusions:

• The proposed method effectively resolves multi-label coupling issues in chest X-ray classification.
• This approach provides a reliable tool for multi-disease collaborative diagnosis, enhancing accuracy in identifying complex lesions.