GraphXCOVID: Explainable deep graph diffusion pseudo-Labelling for identifying COVID-19 on chest X-rays.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Science

Background:

• The COVID-19 pandemic spurred development of automated diagnostic tools using chest X-rays.
• Deep supervised learning models require extensive labeled datasets, which are costly and time-consuming to create for novel diseases like COVID-19.
• Semi-supervised learning offers a promising alternative, achieving high performance with minimal labeled data.

Purpose of the Study:

• To develop and evaluate a novel graph-based deep semi-supervised framework for COVID-19 classification from chest X-rays.
• To demonstrate the efficacy of the proposed framework in outperforming existing supervised models with limited labeled data.
• To enhance diagnostic interpretability for radiologists through attention maps.

Main Methods:

• A graph-based deep semi-supervised framework was developed for COVID-19 classification.
• An optimization model for graph diffusion was introduced to leverage relationships between labeled and unlabeled data.
• Diffusion predictions were used as pseudo-labels in an iterative deep network training scheme.

Main Results:

• The proposed semi-supervised framework outperformed leading supervised models in COVID-19 classification.
• The model achieved high performance with a significantly smaller fraction of labeled examples compared to supervised methods.
• Attention maps were generated to aid radiologists in diagnostic assessment.

Conclusions:

• Graph-based semi-supervised learning is highly effective for COVID-19 diagnosis from chest X-rays, even with minimal labeled data.
• The framework offers a more efficient and potentially more accurate approach to automated disease detection.
• Visualizations enhance model interpretability and support clinical decision-making for radiologists.