An Improved Algorithm for Chest X-Ray Image Classification

Abstract

The application of a Deep Neural Network for medical image classification has been known widely to help the tasks of an accurate doctor assessment. One of the applications is the classification of diseases from Chest X-Ray images. The NIH Chest X-Ray dataset is the most popular and the most prominent medical images database in the field. Several approaches have been made to improve the classification of the 14 classes from the dataset, including modifying network layers, handcrafting the dataset, and conducting smart-augmentation. We propose a weight modification algorithm to overcome the class imbalance problem. Hence, our objective is to improve the classification. The work's novelty is the proposed framework to improve the classification performance, which includes: (i) a novel weight calculation formula, (ii) the real-dataset augmentation into the training data. Experimental results are provided to show the improved classification performance. Our experiments are based on the standardized split-sets, which are also used by previous researches. The steady-state experiment settings are required to achieve acceptable results for the dataset. This study contributes to (i) the further cost-sensitive algorithm to train an imbalance Chest X-Ray dataset, also (ii) we provide results under identical settings compared to the previous study.