A Modification of SynthEnsemble: A Different Approach to the Fusion of Diverse Classifiers for Automated Thoracic Disease Diagnosis

Accumulated Chest X-Ray images have widely been used in the task of automating the diagnosis of thoracic diseases. Multiple studies on using single classifiers for this task and also some studies on the fusion of classifiers for the same purpose have been previously done and all have reached promising results. On furthering the classifier fusion approach, we have implemented a variation of methods including those used in the SynthEnsemble [1] paper on the ChestX-ray14 dataset [2], while focusing on the balance of classifier diversity and accuracy and using the same pre-trained single classifiers from SynthEnsemble as grounds for comparison. We have conducted all the same decision fusion methods, such as weighted averaging and using stochastic global optimization methods for classifier weight optimization, and have also employed new and additional methods, such as various heuristic and intelligent weight optimization methods, with regards to optimizing the number of base classifiers for reducing the computational cost and tuning the accumulated diversity of the fused network. The proposed approach, in summary, is a comparison between a brute-force based global heuristic search as a classifier weight optimizer, and genetically inspired stochastic weight optimizations added to an ordered weighted averaging fusion method, complemented with minimizing the set of base classifiers in accordance to their accuracy-diversity trade-off. Throughout this study, we use various metrics of model evaluation through all of which we have reached better or at least similar performance results while lowering the diversity of the set of chosen classifiers, reducing the number of base classifiers from six to four, and decreasing the execution time by a notable $98.7 \%$. Even notwithstanding the improved performance of our proposed fusion methods, which is greatly dependent on the choice of base classifiers and is strictly limited to those used in [1] for accurate comparison, we can draw the general conclusion of our model’s superiority due to its much smaller execution time, and how it can be of advantage when dealing with real life cases of patient diagnosis, or under the circumstances of employing deeper and more complex hybrid base classifiers for decision fusion.