Enhancing medical image classification through controlled diversity in ensemble learning

Ensemble models in classification problems often encounter learning collision during joint training, where multiple base learners learn similar data representations concurrently. This phenomenon can diminish diversity and confidence in classification, especially in the context of medical image analysis, potentially leading to biased class predictions and classification errors. In this study, we tackle this issue by proposing ensemble models that combine both joint and independent training methodologies on the same medical image dataset. Our key contribution lies in explicitly controlling diversity through the design of the loss function. We finetuned the ResNet50V2 base learner, resulting in a significant 3% increase in training accuracy (86.00% compared to the previous 83.00%). Losses decreased from 0.42 +/- 0.44 to 0.40 +/- 0.42. For ResNet101V2, we observed a 2.27 percentage point increase in training accuracy (84.57% compared to the previous 82.30%) and reduced loss values to 0.357 +/- 0.367 from the previous 0.428 +/- 0.448. Furthermore, we conducted a comparative analysis of our optimized ensemble models, both with and without pruning, to assess their impact on model performance and to better understand their efficacy compared to earlier research work. The results underscore the effectiveness of our approach in mitigating learning collision and enhancing classification accuracy, particularly in the domain of medical image classification. Overall, our approach effectively reduces learning collision and improves classification accuracy as well as test accuracy on unseen medical images, addressing a significant gap in COVID-19 identification. This novel approach holds promise for ensemble models in medical image classification, particularly for lung-related diseases.