Pulmonary Lesion Classification from Endobronchial Ultrasonogram using Adaptive Ray Tracing

Abstract

According to statistical data, lung cancer yields the highest mortality rate. Prior diagnosis of tumor is suggested to lower the rate, since the sooner the lesion is discovered, the more likely it has not developed to be a cancer yet. Benign and malignant lesion identification is still a doubtful field of study to be worked on. Once endobronchial ultrasound (EBUS)—a visualizing tool for an internal bronchus when performing bronchoscopy, became popular due to its advantages, an image obtained from EBUS is also interesting since it is able to reveal the histopathology characteristics of lesion by observing the image. As a result, the classification system based on EBUS image is developed and EBUS data of 96 patients are experimented in this research. The proposed system includes preprocessing on individual EBUS images, detection of lesion boundary, representative frame selection, echoic and statistical feature extraction, and classification. From the experiments, the overall results are quite promising with 68.11% average cross-validation test accuracy and 90% test accuracy of the best model.