Desktop version

Home arrow Health


Anatomy The pancreas is located in the lesser sac deep in the upper abdomen. It runs horizontally from liver to spleen, as shown in Fig. 3.76, and its shape and location differ greatly between subjects. It has a specific relationship with the surrounding vessels. The splenic vein runs along its length, and the superior mesenteric vein runs posterior and perpendicular to the pancreas as it joins the portal vein. All these vessels enhance with contrast (see Fig. 3.77).

Overview of Pancreas Segmentation Algorithms Pancreatic segmentation from a CT volume is crucial for subsequent detailed image analysis of pancreatic pathology in a CAD system. So far, several algorithms have been proposed. One study [211] used an algorithm that analyzed contrast-enhanced images; however, part of the algorithm was executed manually, and it was based primarily on two-dimensional image processing, which can degrade segmentation accuracy. A more sophisticated automated pancreas segmentation algorithm has been presented [158], but it also suffers from low segmentation accuracy due to large variations in the location and shape of the pancreas. A third report [255] presented an algorithm that extracted the pancreas as well as 11 surrounding organs simultaneously from a non-contrast CT volume. A multi-atlas-based algorithm was proposed to extract multiple abdominal

Fig. 3.76

Anatomy of neighboring organs and vessels closely associated with the pancreas

Examples of input contrast-enhanced CT volumes of the pancreas segmentation algorithm (a) Early phase, (b) portal venous phase, (c) late phase

Fig. 3.77 Examples of input contrast-enhanced CT volumes of the pancreas segmentation algorithm (a) Early phase, (b) portal venous phase, (c) late phase

organs including pancreas [313]. Although multi-organ segmentation algorithms might be useful for pancreas segmentation, such algorithms are explained in Sect. 3.9.6. This section focuses on a single-organ segmentation algorithm, or an automated pancreas segmentation algorithm uses a CA model, or an SSM, from contrast-enhanced multiphase CT volumes [258] [259].

CA Models of the Pancreas and Their Application to Segmentation from CT Images The inputs are three-phase CT volume data: early/arterial-, portal-, and venous-phase volumes are presented in Fig. 3.77. Once the three-phase volumes are aligned by a registration algorithm based on normalized mutual information and radial basis function, segmentation of the liver and spleen is performed. The segmentation process assigns a label to each voxel, based on MAP using a probabilistic atlas and parameters estimated by an expectation maximization algorithm. The portal, splenic, and superior mesenteric veins are then extracted by a region-growing-based algorithm using location information for the extracted liver and spleen to establish landmarks for pancreas registration between different subjects. An input patient volume is subsequently warped in a nonlinear fashion with a radial basis function, such that landmarks in the input volume coincide with those in the reference volume data. The second stage roughly extracts the pancreas in the warped CT volume. This is based on the MAP method with a patient-specific probabilistic atlas generated from an SSM of the pancreas or a level set distribution model. After this rough segmentation, a morphological operation with a classifier ensemble is performed to refine the boundary further.

The algorithm was trained using three-phase CT volumes from 98 cases whose size was 512 x 512 x 161-261 voxels at a section interval of 1 mm. The pixel interval in the axial direction ranged from 0.546 to 0.625 mm. To validate the performance of the algorithm, it was applied to three-phase CT volumes from 20 test cases. Figure 3.78 shows examples of the segmentation for a test case where the JI between the extracted pancreas and the true one was 0.560 using a MAP segmentation step and 0.699 using a fine segmentation step. In summary, the average JI of 20 cases was 0.579 using the fine segmentation step.

Compared with the segmentation performance in other organs, such as the liver (see Sect. 3.9.1), the JI is relatively low. One possible reason is that since the pancreas has a slightly lobulated texture, is closely applied to surrounding structures, and is small in volume, the interobserver variability of segmentation is large. In fact, the JI between two true pancreas regions manually delineated by two independent observers (computer engineers) who have carried out studies on medical image processing was 0.760 on average, which is much lower than that for liver segmentation. Although the low JI of pancreas segmentation may be explained by interobserver variability, it does not fully explain some differences between the

Examples of MAP segmentation, fine segmentation, and its corresponding ground truth

Fig. 3.78 Examples of MAP segmentation, fine segmentation, and its corresponding ground truth. (a) Rough segment on result, (b) fine segmentation result, (c) ground truth segmentation performance and the JI. Variability in location, shape, and CT value of the pancreas, as well as essential difficulties in segmentation of a deeply indented and lobulated surface, are also contributing factors.

Computer-Aided Diagnosis in the Pancreas Pancreatic cancer is one of the major causes of cancer-related mortality in Japan, accounting for 29,916 deaths in 2012 [55]. CT is the most widely used imaging examination for detection and staging. CAD systems would be helpful to support the radiological interpretations and findings.

A few algorithms for diagnosis of pancreatic lesions have been presented. One group [130] presented an algorithm to discriminate between pancreatic ductal adenocarcinoma and mass-forming pancreatitis based on the radiological findings extracted by a radiologist, in which no medical image processing was performed. A fuzzy c-means clustering-based tumor extraction algorithm was presented by a second group [144], in which a tumor region was identified by combining the clustering result with a simple manual input by a user. The algorithm was applied to a few data, and the results were validated visually. A quantitative validation study using a large database will be an important future goal in this field.

< Prev   CONTENTS   Source   Next >

Related topics