Deep Learning COVID-19 Features on CXR using Limited Training Data Sets (2004.05758v2)

Abstract: Under the global pandemic of COVID-19, the use of artificial intelligence to analyze chest X-ray (CXR) image for COVID-19 diagnosis and patient triage is becoming important. Unfortunately, due to the emergent nature of the COVID-19 pandemic, a systematic collection of the CXR data set for deep neural network training is difficult. To address this problem, here we propose a patch-based convolutional neural network approach with a relatively small number of trainable parameters for COVID-19 diagnosis. The proposed method is inspired by our statistical analysis of the potential imaging biomarkers of the CXR radiographs. Experimental results show that our method achieves state-of-the-art performance and provides clinically interpretable saliency maps, which are useful for COVID-19 diagnosis and patient triage.

Deep Learning COVID-19 Features on CXR Using Limited Training Datasets

The paper presents a novel deep learning methodology for COVID-19 diagnosis using chest X-ray (CXR) images. Due to the impracticality of collecting large, well-curated datasets under the circumstances of the COVID-19 pandemic, the authors introduce a patch-based convolutional neural network (CNN) trained with a limited dataset. The approach leverages CXR images to identify COVID-19-induced pneumonia, focusing on addressing the need for efficient diagnostic tools given the immense strain on global healthcare systems.

Methodology and Network Architecture

The authors propose a patch-based learning method for constructing CNNs specifically tailored for COVID-19 diagnosis. This method focuses on random patch cropping and utilizes majority voting for classification, thus reducing the overall complexity of the network and mitigating overfitting—an issue particularly concerning due to the limited dataset. The network comprises both segmentation and classification modules, using a fully convolutional DenseNet segmentation network and a simplified ResNet-18 for classification. The presented architecture offers a potential framework for training effective models with limited data by enhancing the training data volume through patch augmentation.

Key Findings

The network outperforms existing methods, often with fewer parameters, making it computationally efficient. It achieves state-of-the-art performance, with a reported sensitivity of 92.5% for distinguishing COVID-19 and viral pneumonia. This statistic is noteworthy, especially given that CXR diagnostic sensitivity is generally lower than RT-PCR tests.

Furthermore, the addition of a probabilistic Grad-CAM—a class activation mapping method—provides interpretability to the model outputs, correlating well with observed radiological findings. This interpretability allows clinical users to visualize potential diagnostic biomarkers on CXRs, contributing to more reliable medical decision-making.

Implications and Future Directions

The findings hold significant implications for the implementation of AI in clinical settings, particularly amidst a pandemic. The ability to effectively utilize limited datasets could transform response capabilities in similar future scenarios, enhancing early-detection systems, resource management, and patient triage strategies. Pragmatically, the proposed method provides a potential triage tool, which could prioritize patients needing urgent care, reserving advanced diagnostic resources for suspected COVID-19 cases.

From a theoretical standpoint, this research exemplifies the utility of patch-based methodologies in reducing overfitting with small datasets and could be expanded to other image-based diagnostic AI systems. Future work might explore enhancing these approaches by integrating other imaging modalities or by further refining architectural elements to increase model robustness and accuracy.

Overall, the paper contributes valuable insights into deep learning applications in medical imaging, particularly under constraining scenarios such as a global pandemic. The patch-based CNN demonstrated in this paper not only underscores the potential of AI to supplement traditional diagnostic methods but also highlights an adaptable framework that can be expanded upon with advancing AI technologies.