Abnormal Object Detection In Thoracic X-Ray Using You Only Look Once (YOLO)
A model to detect abnormal objects using the Yolov5 algorithm on thorax X-ray images was constructed using several methods to improve model accuracy: weighted boxes fusion, image transformation using Contrast Limited Adaptive Histogram (CLAHE), and data augmentation.
Abstract
In diagnosing pulmonary diseases, physicians usually perform a radiologic examination on pulmonary conditions by ensuring particular findings or abnormal conditions. Detecting abnormal conditions may be conducted using a thorax X-ray. This study aimed to construct a model to detect abnormal objects using the Yolov5 algorithm on thorax X-ray images. Fourteen abnormal objects were detected: aortic enlargement, atelectasis, calcification, cardiomegaly, consolidation, ILD (interstitial lung disease), infiltration, lung opacity, nodule/mass, and other lesions, pleural effusion, pleural thickening, pneumothorax, and pulmonary fibrosis. The training model was built using several methods to improve model accuracy: weighted boxes fusion (WBF), image transformation using Contrast Limited Adaptive Histogram (CLAHE), and data augmentation. Experiments on the dataset thorax x-ray showed 61% accuracy of the training model. The trials were performed using different processed datasets; hence, dataset variation and amount play a vital role in improving accuracy. Future studies should implement the model ensemble concept by combining several algorithm models to improve the prediction result.