This project is a collaboration with researches and developers at Memorial Sloan Kettering Cancer Center (MSKCC) who support radiation treatment planning. A crucial part of radiation treatment planning is optimizing how to administer the radiation, position and shape the beams, and adjust the beam intensities so that the clinically prescribed dose is delivered to the treatment area but minimal radiation is applied to surrounding healthy tissue. This requires a volumetric “map” of the body delineating the normal organs and tumor areas.
In this paper we describe the framework we developed to address the problem of segmentation of volumetric medical images. While this paper is restricted to normal tissue and grayscale CT images, we have subsequently generalized the work to handle multi-modal images and tumor tissue. We also have adapted our results from bi-class segmentation to multi-class segmentation. The latest results have been accepted subject to revision, but this paper outlines the basic framework.
Current clinical practice requires that an expert-such as a medical physicist or a physician-manually contour the structures in each slice of a volumetric medical image. Due to the labor intensive nature of this work, it can be quite time consuming and result in fatigue. The results also exhibit differences in judgment from expert to expert, which results in inconsistent results. While many automatic algorithms have been proposed to address this problem, entirely removing a human expert from the process raises ethical and legal issues when the algorithm produces errors. Also, the optimal parameters for an automatic technique may change from device to device, or patient to patient. Another option is to allow for an automated segmentation followed by human correction, but this may often not be any faster than manual segmentation (particularly for challenging anatomic structures).
The principle idea of this approach is to build a statistical model of the structures segmented by the expert online and then progressively improve the model as the segmentation proceeds. Thus we integrate the training which semi-automates the segmentation with a correction step which provides new information to improve the segmentation. In addition, rather than requiring the expert to carefully segment the boundary, the expert simply makes rough brush strokes providing samples within the slice of the structure to be segmented, and its complement. Using this input, a conditional random field is trained based on the statistics of the data under the user brush strokes. Based on these brush strokes, a best estimate segmentation is produced using a graph-cut-based algorithm for conditional random fields. The resulting segmentation is presented to the user for review; the user then accepts or corrects the segmentation. Once the complete segmentation of the slice is accepted, statistics for the region and joint statistics of neighboring pixels are recorded. These further inform the model and improve the results. A proposed segmentation is then available for review in all subsequent slices.
Evaluation of this work is challenging because, given it is a semi-automatic method, it is not meant to directly compete with fully automatic methods. Since an expert remains part of the processes, it is always possible to maintain human performance levels through human intervention. One method of evaluation applies the algorithm semi-automatically in one slice, but treats the proposed segmentation in the next slice as automatic and compares it with baseline automatic techniques. Another evaluation technique is to compare the improvement in speed.