Deep generative abnormal lesion emphasization validated by nine radiologists and 1000 chest X-rays with lung nodules.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Accurate detection of abnormal lesions in chest radiographs is crucial for early disease diagnosis.
• Existing methods may have limitations in highlighting subtle abnormalities, impacting diagnostic accuracy.

Purpose of the Study:

• To introduce EGGPALE (Extrapolative, Generative and General-Purpose Abnormal Lesion Emphasizer), a new method for emphasizing abnormal lesions in chest X-rays.
• To evaluate the effectiveness of EGGPALE in improving the diagnostic performance of radiologists.

Main Methods:

• EGGPALE utilizes a flow-based generative model trained on normal chest radiographs.
• It employs latent space extrapolation to move image representations away from the normal data distribution.
• The method generates an emphasized image by mapping the extrapolated latent representation back to the image space.

Main Results:

• EGGPALE processing resulted in an average sensitivity improvement of +0.0559 compared to original images.
• A slight average specificity deterioration of -0.0192 was observed.
• The area under the receiver operating characteristic curve showed a statistically significant improvement when using EGGPALE-processed images.

Conclusions:

• EGGPALE demonstrates feasibility and effectiveness in enhancing abnormal lesions in chest radiographs.
• The method shows potential to aid radiologists in improving diagnostic accuracy for conditions like lung cancer.
• Further validation and integration into clinical workflows are warranted.