Desktop version

Home arrow Health

Multi-organ Computational Anatomy Modeling: Organ Correlation Graph

(OCG) To represent interrelations among multiple organs more generally, a graph representation combining the basic units described in the above paragraph is formulated. Figure 3.89 shows a graph representation called the organ correlation graph (OCG). In the OCG, the basic unit is represented by nodes and directed arcs of a graph, in which the nodes correspond to organs and the directed arcs connect the predictor to target organs. One problem is how to find the arcs between

Conventional and prediction-based probabilistic atlas of (a) gallbladder and (b) pancreas. Left

Fig. 3.88 Conventional and prediction-based probabilistic atlas of (a) gallbladder and (b) pancreas. Left: Single-organ modeling. Right: Multi-organ modeling incorporating interrelation with the liver and spleen. P (Pan) denotes the probabilistic atlas of the pancreas while P (PanjLiv, Sp) the probabilistic atlas of the pancreas under the condition that the regions of the liver and spleen are known

Organ correlation graph (OCG) of the upper abdominal organs

Fig. 3.89 Organ correlation graph (OCG) of the upper abdominal organs

nodes. Basically, the arcs are determined so as to minimize prediction error. For example, the predictor organs predicting the gallbladder are selected so that the prediction error is minimized, and the arcs are connected from the selected predictor organ nodes to the gallbladder node. Manually defined constraints can also be incorporated. For example, the liver is regarded as an anchor organ, and its region is assumed to be segmented beforehand using the methods as described in the previous section. That is, the liver can only be a predictor organ node and any arcs do not direct to it. Conversely, the gallbladder can only be a target organ node because its segmentation may not be accurate enough to use as one of the predictor organs. Once the set of organ nodes and the abovementioned manually defined constraints on them are given, the directed arcs in the OCG are determined automatically by finding the incoming arcs to each target organ node, which minimize the prediction error subject to satisfying these manually defined constraints. By using the OCG, the prediction-based CA models (i.e., prediction-based probabilistic atlas and SSM) can be generated once the regions of the predictor organs are segmented.

Multi-organ Segmentation The OCG is applied to automated multi-organ segmentation. The OCG is used for generating a procedure for automated multi-organ segmentation from abdominal CT data. The basic assumptions are as follows: (1) The field of view (FOV) of input CT data includes the whole liver, which is the anchor organ. (2) The intensity models of each organ and its background, which are the probability distributions of the CT values inside and outside the organ region, respectively, are available (or a procedure for estimating the intensity models by using the OCG and input CT data is available). The segmentation method at each node of the OCG is basically the same as that which was applied to liver segmentation (described in the previous section), but the difference is that the conventional probabilistic atlas and SSM are replaced by the prediction-based

S. Hanaoka et al.


ones. Each organ node is prepared to start the segmentation procedure when the segmentation results at the predictor organ nodes are obtained. The segmentation procedures are executed at all the prepared nodes in a synchronized manner in the OCG and repeated several times to obtain the final segmentation results at all the nodes.

The abovementioned multi-organ segmentation method was tested using more than 100 CT datasets obtained under four different imaging conditions in contrast agent and CT scanner at two hospitals. The intensity models were constructed for each imaging protocol while the same priors on shape and location were utilized for all the datasets. Leave-one-out cross validation was performed. Figure 3.90 shows typical results. The prediction-based priors were effective in these results. JI was used for accuracy evaluation.

Results of abdominal multi-organ segmentation from CT data

Fig. 3.90 Results of abdominal multi-organ segmentation from CT data. Table shows JI of results of multi-organ (prediction-based) and single-organ (conventional) methods. The improvement of segmentation accuracy was notable in the organs surrounded by red rectangles in the Table

< Prev   CONTENTS   Source   Next >

Related topics