An interactive color pre-processing method to improve tumor segmentation in digital medical images
Is Version Of
In the last few decades the medical imaging field has grown considerably, and new techniques such as computerized axial tomography (CAT) and Magnetic Resonance Imaging (MRI) are able to obtain medical images in noninvasive ways. These new technologies have opened the medical field, offering opportunities to improve patient diagnosis, education and training, treatment monitoring, and surgery planning. One of these opportunities is in the tumor segmentation field.
Tumor segmentation is the process of virtually extracting the tumor from the healthy tissues of the body by computer algorithms. This is a complex process since tumors have different shapes, sizes, tissue densities, and locations. The algorithms that have been developed cannot take into account all these variations and higher accuracy is achieved with specialized methods that generally work with specific types of tissue data.
In this thesis a color pre-processing method for segmentation is presented. Most tumor segmentation methods are based on grayscale values of the medical images. The method proposed in this thesis adds color information to the original values of the image. The user selects the region of interest (ROI), usually the tumor, from the grayscale medical image and from this initial selection, the image is mapped into a colored space. Tissue densities that are part of the tumor are assigned an RGB component and any tissues outside the tumor are set to black. The user can tweak the color ranges in real time to achieve better results, in cases where the tumor pixels are non-homogenous in terms of intensity. The user then places a seed in the center of the tumor and begins segmentation. A pixel in the image is segmented as part of the tumor if it's within an initial 10% threshold. This threshold is determined if the seed is within the average RGB values of the tumor, and within the search region. The search region is calculated by growing or shrinking the previous region using the information or previous segmented regions of the set of slices. The method automatically segments all the slices on the set from the inputs of the first slice. All through the segmentation process the user can tweak different parameters and visualize the segmentation results in real time.
The method was run on ten test cases several runs were performed for each test cases. 10 out of the 20 test runs gave false positives of 25% or less, and 10 out of the 20 test runs gave false negatives of 25% or less. Using only grayscale thresholding methods the results for the same test cases show a false positive of up to 52% on the easy cases and up to 284% on the difficult cases, and false negatives of up to 14% on the easy cases and up to 99% on the difficult cases. While the results of the grayscale and color pre-processing methods on easy cases were similar, the results of color pre-processing were much better on difficult cases, thus supporting the claim that adding color to medical images for segmentation can significantly improve accuracy of tumor segmentation.