| Abstract |
Intensive research has been made during the last decades in the field of modeling of tumor growth and response to different therapeutic modalities. The ultimate goal of simulating cancer is to help to better understand cancer biology and substantially improve treatment planning, by fully exploiting the individual data of each patient as well as predicting the response to potential therapeutical schemes. The incorporation of imaging data into the simulation process is of extreme importance, not only for individualizing the procedure, but also for the clinical evaluation of the predicted outcome. However, tumors appear in medical images as inherently inhomogeneous structures, with sometimes poorly defined, jagged edges, a fact that makes the tumor segmentation process a difficult task. Also, in many cases, tumor images contain small, sharp-edged, inhomogeneous features, which could be important cancer necrotic areas or cysts, but difficult to be included in the segmentation result and for this reason clinically interactive segmentation is usually recommended.
In order to overcome the aforementioned difficulties, we propose a semi-automatic method for interactive tumor segmentation, based on traditional active contours (or snakes). Snakes are flexible curves that are based on the deformation of an initial contour towards the boundary of the desired object, which is accomplished by minimizing a suitable energy functional, designed so that its minimum is obtained at the desired boundary. This functional consists of the summation of three energy terms, one controls the smoothness of the curve and the second attracts the curve towards the desired boundary, while the last one is a pressure force, pushing the snake to deform outwards, just like an inflating balloon. Snake evolution can be driven by weighting parameters which control the impact of each energy term on the resulting contour.
In this work, we improved traditional active contours by introducing local snake bending. The key point of our method is the use of adaptable parameters, instead of constant ones, for the snake evolution. Considering that a snake, during its deformation, should be rigid and forceful inside the tumor, in order to avoid getting caught by small sharp-edged regions, and very flexible near the true borders, in order to accurately adjust to boundary details, we group image pixels into two groups, according to underlying gradient magnitude and corner strength characteristics, and assign to each group different sets of parameters. Thus, the improved algorithm is able to spatially adapt the snake's behavior to image features and include, or not, small high-contrast regions, according to their gradient and curvature characteristics, while, at the same time, it can accurately detect boundary details. Furthermore, we developed an additional easy-to-use feature for interactive refinement of the obtained result, in order to facilitate the tiresome and error-prone manual result correction.
The method is applied on 157 MR images of renal tumors and then compared to the traditional snake method and the region growing technique. Using the clinical expert’s annotations of the true tumor boundaries in all 157 images as gold standard, we demonstrate that our method accurately delineates tumor borders, even in difficult cases of bad image quality.
|
Search
