Radiologist-supervised Transfer Learning: Improving Radiographic Localization of Pneumonia and Prognostication of...

Area of Science:

• Medical imaging analysis
• Artificial intelligence in radiology
• Transfer learning applications

Background:

• Convolutional neural networks (CNNs) show promise for pneumonia detection.
• Radiographic severity assessment is crucial for predicting COVID-19 patient outcomes.
• Enhancing CNN performance with expert input is an active research area.

Purpose of the Study:

• To evaluate radiologist-supervised transfer learning for improving CNN-based pneumonia localization.
• To assess the prognostic capability of CNN-derived pneumonia severity quantification in COVID-19 patients.

Main Methods:

• A pre-trained CNN was fine-tuned using transfer learning with radiologist-annotated radiographs.
• Performance was evaluated using area under the curve (AUC) and Dice similarity coefficient.
• Prognostic value was assessed via survival analysis in COVID-19 patients, compared to the mRALE score.

Main Results:

• Pneumonia detection AUC improved significantly after fine-tuning.
• Dice overlap demonstrated enhanced localization, especially in lung bases.
• CNN severity quantification showed strong correlation with radiologist scores and similar prognostic ability for mortality and intubation risk.

Conclusions:

• Radiologist-supervised transfer learning effectively enhances CNN performance for pneumonia localization and quantification.
• Integrating radiologists into AI development offers a closed-loop system for continuous algorithm improvement.