Assessment of mass classification from chest X-ray using double-transfer learning: a comparative study with diverse...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Chest X-ray analysis is critical for diagnosing various lung diseases.
• Artificial intelligence (AI) models are increasingly used for medical image classification.
• Transfer learning (TL) techniques can enhance AI model performance by leveraging pre-trained models.

Purpose of the Study:

• To evaluate the effectiveness of double-transfer learning (double-TL) in improving AI model performance for mass classification in chest X-ray images.
• To compare the performance of AI models trained with double-TL against those trained with single-transfer learning (single-TL).

Main Methods:

• Chest X-ray images depicting infiltration, atelectasis, effusion, and nodules were utilized for the single-TL phase of the double-TL process.
• Four AI models were trained using double-TL: Infiltration-Mass, Atelectasis-Mass, Effusion-Mass, and Nodule-Mass.
• A comparative AI model, Mass, was trained using single-TL.

Main Results:

• Area under the curve (AUC) values for double-TL models ranged from 0.783 to 0.795, comparable to the single-TL Mass model's AUC of 0.788.
• The Nodule-Mass model achieved the highest recall for disease (0.75) and precision for healthy cases (0.86).
• Overall, double-TL models demonstrated comparable or marginally improved performance compared to the single-TL Mass model.

Conclusions:

• Double-transfer learning (double-TL) yielded marginal improvements in specific models, like enhanced sensitivity in the Nodule-Mass model, but overall performance gains were limited.
• Double-TL may offer benefits for particular tasks but did not consistently surpass single-TL performance.
• The efficacy of double-TL is significantly influenced by the size and feature similarity of selected images, underscoring the importance of strategic image selection in healthcare AI.