Lung nodules are masses of tissue in the lung which on a chest X-ray or computerized tomography (CT) scan appear as round, white shadows. Though lung nodules are quite common and most of them are noncancerous (benign), some nodules may be cancerous and it is extremely important to identify them early in their course, when they can still be effectively treated and cured. In addition, segmentation of lung tumors helps finding out how far they have spread (staging), their response to treatments as well as their progression in size and shape.
Medical experts can visually identify nodules on a CT scan, but they need to rely on computerized analysis to estimate shape and volume of nodules, as well as their progression over time. It is visually easier to identify nodules than lung tumors, since the former are supposed to have an elliptical shape, while the chromatic aspect of the latter is difficult to distinguish from healthy tissues on a CT image.
Our client needed an automatic process of lung nodules segmentation to deal with this difficulty in a way that is quick to perform and memory efficient. Our methodology is built around graph cut algorithms: graph cut is a technique in the field of computer vision (borrowed from the world of graphics) which offers very efficient solutions to a wide range of problems in vision and graphics. In particular, it provides a simple and effective answer to problems of segmentation of objects in image data. Its best features include great robustness and practical efficiency when facing the constraints of multidimensional analysis. We at RSIP Vision have found graph cuts segmentation in its tridimensional application the best approach to perform nodule segmentation in lung CT images.
The graph cut technique gives us a comfortable tool for user interaction, i.e. how the user and the system act on each other. This is particularly valuable with regard to 3D segmentation. The uniqueness of the graph cut segmentation approach and of our method to implement it in lung nodule segmentation is that the technique is built around a concept of binary energy minimization. The purpose is to specify whether a pixel is positioned inside or outside the object of interest. Energy is minimized by imposing smoothness constraints on the shape of the desired object, in such a way that its edges are highlighted and a solution is found which is as close as possible to the true form of the studied object.
To obtain satisfactory results, these models require the presence of a gradient in the nodule being reviewed which is different from the one of the background. We have found that the most effective way of dealing with this issue is having the expert define the input data: an object seed which will be confronted with the background seed starting from an initial point, typically selected in the center of a nodule.
Compared with hand-made nodule segmentation in the lung (performed by an expert), this automatic lung nodules segmentation method grants results which are very close to the (much slower) manual segmentation. It also enables our client to overcome the difficulty of tumors and healthy tissues appearing on the CT scan in a very similar range of intensity values and it allows a skilled expert to position the initial seed in such a way that the calculation of minimal energy for the segmentation of lung tumors will take only a few seconds. You can read here about other projects led by RSIP Vision in the field of pulmonology.