Optimizing Medical Image Diagnosis: The Role of Feature Selection in Deep Learning Models

The disorders that affect the lungs require a careful diagnosis, which relies on medical imaging such as chest X-rays for efficient detection. Convolutional neural networks show a lot of promise as a deep learning model in the automation of medical diagnosis, though in most cases, their success ends up being attached to how well the right features are chosen. In this work, three broad types of feature selection methods, which are the filter methods, the wrapper methods and the embedded methods, are analyzed on a chest X-ray dataset for their role in improving the performance of deep-learning methods in the classification of medical images. The evaluation metrics of the individual methods, which are classified into their specific broad types, are calculated. The results show that embedded techniques like Light Gradient Boosting Machine, along with filter methods such as Independent Component Analysis and Recursive Feature Elimination, which is a wrapper method, performed better, while the Mutual Information, which is a filter method, performed worst among the methods. The results eventually showed the highest difference in accuracy, being about 7.9%. In conclusion, these results establish the key role that the correct choice of a feature selection method plays in efficient medical image classification.