Effect of tumor delineation strategies on ANN classification accuracy in lung CAD

Lung Cancer is a serious illness and patient survival rate depends on early and accurate detection. CAD systems are commonly used for detection and characterization of nodules. The type of tumor segmentation algorithm or radiologist segmentation may affect the accuracy when characterizing lung nodules on chest x-ray images. In order to segment and classify nodules better, preprocessing step is needed. Histogram equalization, fuzzy minimization, bone subtraction, cropping can be some steps of preprocessing. In this study, the main object is to evaluate the accuracy of the characterization of lung nodules on bone subtracted chest x-ray images by using different types of boundary segmentation algorithms and an artificial neural network based classification method. Another aim is to evaluate the contribution of CAD systems and accuracy of radiologist segmentation on raw chest X-ray. The standard digital image database with chest lung nodules (JSRT database) that was created by the Japanese Society of Radiological Technology in cooperation with the Japanese Radiological Society (JRS) is used. To subtract the bones a bone shadow elimination algorithm is used. Preprocessing and look up tables are used if nodule is not clearly seen. Active contour, spline active contour and radiologist based delineation methods are used. Artificial neural network classifications are used and their accuracy is evaluated. At the end, high specificity and sensitivity ratios are obtained and different segmentation techniques are compared. As a result, results are satisfying and interesting. Future work is possible to extend the study to other segmentation techniques and modalities.