Image processing techniques for lung lesion detection in chest CT scans

Abstract

Imaging is an essential component in modern medicine, especially in areas that deal with nonvisible internal organs, such as the lung. Lung lesion detection is an important step in the diagnosis of several diseases, including lung cancer, which is considered to be one of the deadliest cancers. With that in mind, Computer Aided Detection (CAD) tools are becoming more popular in medical diagnosis. Such tools aim to support practitioners in their jobs by reducing human error as much as possible. Image processing is the backbone of CAD systems. It is used by such systems to enhance the quality of the image, segment the region of interest, highlight features within the image, and support anomaly detection techniques. Computed Tomography (CT) is an X-Ray based medical imaging modality which generates hundreds of images per patient. Therefore, developing CAD systems for this particular type of imaging modality is crucial. This study applies image processing techniques on CT Scans of the thorax, designs and implements an automatic CAD tool aimed at detecting lesions in lungs. This is done by pre-processing the CT images to remove noise through the use of a Bilateral Filter and enhance the contrast by using Gamma Correction and Contrast Limited Adaptive Histogram Equalisation. In both these pre-processing steps, the parameters required for the techniques implemented, were verified by an expert radiologist to ensure that no anatomical details are removed by these techniques. The lungs are segmented by means of a marker-based watershed algorithm. Detection is performed by applying a priori knowledge about lesion intensities and shapes. Lesion characteristics are used by the CAD system developed to accurately separate lesions from vessels. This helps in reducing false positives, that is, the detection of a lesion in the absence of one. The CAD system designed was tested on 25 cases, achieving a sensitivity of 76.3%, an average false positive rate of 7.88 lesions per scan and an average false negative rate of 0.56 lesions per scan. This is done in an average time of 74.3ms per slice.