Ensemble Learning based on Classifier Prediction Confidence and Comprehensive Learning Particle Swarm Optimisation for Medical Image Segmentation

Abstract

Segmentation, a process of partitioning an image into multiple segments to locate objects and boundaries, is considered one of the most essential medical imaging process. In recent years, Deep Neural Networks (DNN) have achieved many notable successes in medical image analysis, including image segmentation. Due to the fact that medical imaging applications require robust, reliable results, it is necessary to devise effective DNN models for medical applications. One solution is to combine multiple DNN models in an ensemble system to obtain better results than using each single DNN model. Ensemble learning is a popular machine learning technique in which multiple models are combined to improve the final results and has been widely used in medical image analysis. In this paper, we propose to measure the confidence in the prediction of each model in the ensemble system and then use an associate threshold to determine whether the confidence is acceptable or not. A segmentation model is selected based on the comparison between the confidence and its associated threshold. The optimal threshold for each segmentation model is found by using Comprehensive Learning Particle Swarm Optimisation (CLPSO), a swarm intelligence algorithm. The Dice coefficient, a popular performance metric for image segmentation, is used as the fitness criteria. The experimental results on three medical image segmentation datasets confirm that our ensemble achieves better results compared to some well-known segmentation models.