The Computer-Aided Diagnosis Of Pneumoconiosis

Pneumoconiosis is the most serious occupational diseases in China. It is reported that in coal industry the number of new pneumoconiosis patients announced have been up to nearly 57 thousands in recent years. The new cases of pneumoconiosis account for about 90% in total cases of occupational diseases. So it is important to develop a computer-aided diagnosis of pneumoconiosis to assist radiologists.The major finding of pneumoconiosis in radiology is nodular opacities in various shapes and sizes. At present, the high-resolution CT scan of thorax is a dominant diagnostic approach. Comparing to chest X-ray images, the pneumoconiosis can clearly manifest itself on volumes data, not be covered by muscle or bones. However, the high resolution of CT scan also means heavy workload for radiologists. In addition, the diagnosis of pneumoconiosis mainly depends on of radiologists' experience. It may lead to misdiagnosis.Our research aims to develop a computer-aided diagnosis system to assist radiologists to make a diagnosis of pneumoconiosis. At first we designed two enhancement filters based on Hessian matrix to extract nodules and vessels respectively. Then we eliminated false positives by subtracting vessels from nodules and checking nodules based on geometric features. Finally the types of pneumoconiosis are classified by SVM classifier. Due to complex local objects in thorax and lung lesions caused by pneumoconiosis, many vessels and bronchus are mistaken for nodules which serious reduce the accuracy of classification of pneumoconiosis cases. In this research, first we focused on the methods of reducing false-positives. Second, we paid attention to select more suitable features of classification according to the diagnosis criterion of pneumoconiosis.The system was evaluated on a data set of 139 pneumoconiosis cases. These cases are divided into category 2 and category 3 (the category 3 can be further divided into category 3-i and category 3-ro) based on physicians' diagnosis. When we classified category 2 and category 3 of cases, we obtained total accuracies of 80%. The average total accuracy was reduced to around 60% when we classified all three categories of cases. It is a promising result and finally we point out the direction and methods to improve the performance of classification in future research.