MI-DenseCFNet: Deep Learning for Multimodal Diagnosis of Aureus and Aspergillus Pneumonia

Overview

This study developed and validated MI-DenseCFNet, a multimodal deep learning model combining CT imaging and clinical data to differentiate Staphylococcus aureus pneumonia (SAP) from Aspergillus pneumonia (ASP). The model integrates DenseNet-201 for image feature extraction with a deep neural network for clinical features, improving diagnostic accuracy over traditional methods.

Background

Lower respiratory tract infections, including pneumonia, cause over 2.49 million deaths globally. Differentiating pneumonia pathogens such as SAP and ASP is critical due to differing treatments and prognoses but remains challenging with conventional imaging interpretation. CT scans provide valuable pathological insights, yet human interpretation is prone to error. Artificial intelligence, particularly deep learning, offers automated feature extraction from imaging and clinical data, potentially enabling faster and more accurate pathogen identification.

Data Highlights

A total of 120 patients (60 SAP and 60 ASP) were included after strict inclusion and exclusion criteria. From these patients, 31,259 high-resolution CT images were collected and preprocessed. The MI-DenseCFNet model utilizes DenseNet-201 for image feature extraction and a three-layer deep neural network for clinical feature extraction, fusing these data for classification. Data augmentation and normalization were applied to improve model training.

Key Findings

• MI-DenseCFNet combines CT imaging features and clinical data for multimodal pneumonia diagnosis.
• DenseNet-201 effectively extracts imaging features from high-resolution CT scans.
• A three-layer deep neural network processes clinical symptoms, laboratory results, and imaging features.
• Fusion of image and clinical feature vectors enhances diagnostic accuracy for distinguishing SAP from ASP.
• Specific imaging signs (air crescent, air sacs, halo sign) aid differentiation but require expert interpretation; the model reduces reliance on subjective judgment.
• Early and accurate pathogen identification supports targeted treatment, potentially improving patient outcomes.

Clinical Implications

The MI-DenseCFNet model offers clinicians a robust tool to differentiate SAP and ASP using routinely available CT imaging combined with clinical data, reducing diagnostic delays and errors. This approach can guide timely, pathogen-specific therapy, minimizing empirical treatment risks and improving prognosis. Integration of such AI models into clinical workflows may enhance decision-making in respiratory infections.

Conclusion

MI-DenseCFNet demonstrates the feasibility and benefit of a multimodal deep learning approach for accurate, early diagnosis of pneumonia caused by Staphylococcus aureus and Aspergillus. This model represents a promising advancement in leveraging AI to improve infectious disease diagnostics.

References

1. Global Burden of Disease 2019 Study -- Mortality from Lower Respiratory Tract Infections
2. American Thoracic Society Guidelines -- Diagnostic Criteria for Pneumonia