Cooperative strategy for a dynamic ensemble of classification models in clinical applications: the case of MRI vertebral compression fractures

Purpose: In clinical practice, the constructive consultation among experts improves the reliability of the diagnosis and leads to the definition of the treatment plan for the patient. Aggregation of the different opinions collected by many experts can be performed at the level of patient information, abnormality delineation, or final assessment. Methods: In this study, we present a novel cooperative strategy that exploits the dynamic contribution of the classification models composing the ensemble to make the final class assignment. As a proof of concept, we applied the proposed approach to the assessment of malignant infiltration in 103 vertebral compression fractures in magnetic resonance images. Results: The results obtained with repeated random subsampling and receiver operating characteristic analysis indicate that the cooperative system statistically improved ((Formula presented.)) the classification accuracy of individual modules as well as of that based on the manual segmentation of the fractures provided by the experts. Conclusions: The performances have been also compared with those obtained with those of standard ensemble classification algorithms showing superior results.

Purpose: In clinical practice, the constructive consultation among experts improves the reliability of the diagnosis and leads to the definition of the treatment plan for the patient. Aggregation of the different opinions collected by many experts can be performed at the level of patient information, abnormality delineation, or final assessment. Methods: In this study, we present a novel cooperative strategy that exploits the dynamic contribution of the classification models composing the ensemble to make the final class assignment. As a proof of concept, we applied the proposed approach to the assessment of malignant infiltration in 103 vertebral compression fractures in magnetic resonance images. Results: The results obtained with repeated random subsampling and receiver operating characteristic analysis indicate that the cooperative system statistically improved (p< 0.01) the classification accuracy of individual modules as well as of that based on the manual segmentation of the fractures provided by the experts. Conclusions: The performances have been also compared with those obtained with those of standard ensemble classification algorithms showing superior results.