Assessment of mass classification from chest X-ray using double-transfer learning: a comparative study with diverse image selection for single-transfer learning

The aim of this study was to assess the efficacy of double-transfer learning (double-TL) in enhancing the artificial intelligence (AI) model performance of mass image classification of chest X-Ray images.

Chest X-ray images of various diseases (infiltration, atelectasis, effusion, and nodule) were selected for the single-TL phase in double-TL process. Then four AI models (Infiltration-Mass, Atelectasis-Mass, Effusion-Mass, and Nodule-Mass) were trained using double-TL. For comparison, an AI model of Mass was trained using single-TL.

The area under the curve (AUC) values were 0.795, 0.785, 0.785, and 0.783 for the Effusion-Mass, Infiltration-Mass, Atelectasis-Mass, and Nodule-Mass models, respectively, while the Mass model showed an AUC of 0.788. The Nodule-Mass model achieved the highest recall for disease (0.75) and precision for healthy (0.86). Overall, the double-TL models showed comparable or slightly improved performance over the Mass model.

Although double-TL showed marginal improvements in some models, such as higher sensitivity in the Nodule-Mass model, overall performance gains were limited. These findings highlight that double-TL may benefit specific tasks but did not consistently outperform single-TL. The performance of double-TL is highly contingent upon both the size and feature similarity of the selected images, indicating that appropriate image selection is crucial for maximising the efficacy of double-TL in healthcare applications.