Hot Spot Detection and Segmentation in whole body FDG-PET Images   Haiying Guan, Toshiro Kubota, Xiaolei Huang, Xiang Sean Zhou, Matthew Turk   June 2005 - Sept. 2005   Siemens Medical Solutions

Overview

Details

1.    Introduction

Positron emission tomography (PET) using fluorine-18 deoxyglucose (FDG) is a nuclear medicine medical imaging technique that produces a three dimensional image of functional processes in the body. Tumors in FDG-PET appear as “hot spots” due to increased FDG uptake, and by late 1990s, a large body of literature have clearly shown that FDG-PET imaging is essential and effective for detection, diagnosis and prognosis of tumors. It is also shown applicable to neurological disorders and coronary artery disease. It is expected that the number of FDG-PET images performed in most facilities will exceed all other procedures  performed with radioactive compound in the coming years.

As the amount of PET data generated every day increases, the burden on the radiologists in reading and diagnosing the cases increases. This may result in a negative effect, as the rate of overlooking anomalies can increase due to fatigue, and the risk of over-diagnosis may also rise. Computer-aided detection (CAD) tools have constantly demonstrated effectiveness in supplementing the task of radiologists; they improve the sensitivity of anomaly detection and reduce the interpretation variation among radiologists. Currently, to our knowledge, there are no commercially available CAD tools for FDG-PET images.

2.    Method

2.1  SYSTEM OVERVIEW

Figure 1 shows an overview of our PET-CAD system by organizing the system components with the processing flow. Because the raw intensity of PET depends on many factors and is very unstable, we first convert PET values to Standardized Uptake Values (SUV). Since FDG intakes of different organs are different, for accurate detection and segmentation, it is essential to understand the body context. Thus the next step in the processing flow is to derive a rough body context interpretation using a spatial Hidden Markov Model (HMM). Next, we extract suspicious regions with abnormal FDG intakes. We employ a computationally efficient competition diffusion (CD) algorithm for the task. The CD is sensitive to initialization, and we use the body section information obtained above to introduce different initialization strategies to lung and other organs. Finally, according to the body context knowledge and detection results, we fine-tune the segmentation by a model-seeking region growing algorithm.

Fig. 1. The system diagram

2.2  STANDARDIZED UPTAKE VALUE (SUV)

The big challenge for PET image processing is that the intensity of PET images heavily depends on the clinical and patient factors. In addition, the intensity value is widely distributed from 0 to 32767. The histogram of PET intensity typically shows no peak. Thus, the traditional Gaussian Mixture Model (GMM) is not suitable for modeling the histogram distribution of intensity values of PET images, even after the normalization.

In clinical literature, SUV is widely used by radiologists for normalization across time points, and different patients. It normalizes the measured FDG concentration by decay corrected injection dose per gram body mass. We convert a raw PET value to the SUV using the following equation.

Figure 3 illustrates the difficulties associated with working directly on PET values. We segment a PET volume using three different techniques: fixed thresholding on PET values, adaptive thresholding on PET values, and fixed thresholding on SUV values. In Figure 2, (a) (b) (c) are three test volumes, and (b) and (c) are volumes from the same patient taken at different time point. For each cases, from left to right, the fixed thresholding on PET values (Alg. I), the adaptive thresholding on PET values (Alg. II), and fixed thresholding on SUV values (Alg. III) are shown. PET values show high variations between two volumes and fixed thresholding cannot cope with the variations; case (b) Alg. I was under segmented and case (c) Alg. I was over segmented. The adaptive thresholding failed to handle inter-patient variations of PET values; Case (b) Alg. II and (c) Alg. II are over-segmented, while case (a) Alg II is severely under-segmented. SUV values exhibit stability against both inter and intra-patient variations. As shown in Figure Alg. III, a simple thresholding drawn at 2.5 could effectively and efficiently extract regions of high FDG uptake.

Fig. 2. Comparisons of three segmentation algorithms

2.3  BODY CONTEXT INTERPRETATION

Malignant cells in different organs exhibit different levels of FDG uptake. For example, cancer cells in liver tend to take more FDG than cancer cells in lung, while a normal liver may take as much FDG as cancer cells in lung. Thus, to accurately detect abnormal FDG uptake in a PET volume, it is important that we have a rough anatomical interpretation of the volume. It is extremely difficult to extract anatomical  information from a PET volume, as the PET data does not present detailed structures of the body. One way to accomplish the task is to use a CT volume of the same patient and register it to the PET volume. Many recent PET scanners are combined with a CT scanner, providing a pair of accurately registered PET-CT volumes. It is thus feasible to provide anatomical interpretation of the PET volume by first extracting bones, lungs and other organs from the CT volume and mapping them onto the PET counterpart. The approach, although highly effective, does require a registered CT data, demand more storage, and make the system less portable. Due to these disadvantages, in the present work, we explore a way to achieve the goal with a PET volume alone.

Our body section labeling algorithm is based on two observations. First, the neck curve when the 3D medial axis of the body is projected to the sagittal plane exhibits a distinctive pattern as shown Fig. 3 (a)). Second, the FDG uptake of the normal tissue of the lung is generally less than other tissues such as liver and kidney. Fig. 3(b) shows one example. We compress the 3D body into two feature curves that preserve the above observations. The feature curve of the first observation is a trace of the centroid of the body in  each axial slice. The feature curve of the second observation is, after removing SUVs above 2.5, a trace of the mean intensity in each axial slices (black arrows in Fig. 3).

Fig. 3. HMM training and recognition

We then employ an Hidden Markov Model(HMM) to divide the body into three sections: neck, lung and abdomen. HMM is a powerful parametric model and is feasible to characterize the stochastic processes: the measurable observation process and the immeasurable hidden states. It has been successfully applied to speech recognition and gesture recognition in the temporal space. In our work, we consider a spatial discrete HMM along the head to toes direction, t. First, two HMM models, neck model and lung model, are obtained from the observation patterns of necks and lungs. The training samples of the two feature curves are extracted by manual markings of the lungs. After vector quantization, for the given training samples, the models are obtained by maximizing the likelihood Pr(Observ.|Model) with Baum-Welch algorithm.

In recognition stage, with the models, the locations of a neck and lungs are obtained by searching for the best match point in the whole body feature curves to the models.

To do so, we compute Pr(Model|Observ.) within a sliding window with the forward-backward algorithm and locate the maximum point. In our system, we post-process the probability curve with the spatial integration, and output the locations of the maximum likelihood point as our final neck or lung point. Figure 3(c) illustrates the recognition process for the lung feature.

2.4 DETECTION OF THE VOLUME OF INTEREST

The Competition-diffusion (CD) algorithm is a segmentation algorithm which assigns to each voxel a class label from a set of n classes. It contains two processes: competition and diffusion. The competition process selects the most fitted label and prevents over smoothing. The diffusion process brings a good spatial coherence to the detection map. Figure 4 shows a result of the VOI detection in the PET image of a patient with Bone Metastasis. The left image is a result of applying thresholding at 2.5 to the SUV volume. The middle image is a result of applying CD with uniform initialization. We used the non-lung initialization strategy as described above. The right image is a result of applying CD with section dependent initialization. CD removed some small clusters whose SUV values are not significantly high, and the body section dependent initialization revealed additional hot spots in the ribs.

Fig. 4. VOI detection results

2.5  MODE-SEEKING REGION GROWING

The CD algorithm with a low cut-off threshold produces VOI that have high probabilities of being hot spots. Because this is a conservative step to detect VOI, the volumes tend to be larger than the true extent of the hot spots (see Fig. 5(b)), and several nearby hot spots may be merged in one volume. In order to achieve accurate quantitative volume measurement of each hot spot, we further apply a Mode-Seeking Region Growing segmentation algorithm (MSRG). The first step is to find the intensity modes (or local maximums), which correspond to the primary locations of hot spots, where the FDG uptake is the maximum. Radiologists often use the maximum SUV value of a hot spot to grade tumors, and help to determine the extent of a hot spot. Hence a segmentation algorithm based on the maximum SUV value produces clinically meaningful and accurate segmentation results that are consistent with a doctor’s interpretation. In our mode-seeking step, we start from a seed point as a probe. At the probe point, we find the maximum SUV location within the neighborhood of the point, we then move the probe to the point with the maximum SUV value in the neighborhood. This process continues until the probe reaches an intensity mode in the spatial domain where the probe itself would have the maximum SUV value and stop moving. If we view the SUV image as a 4D surface f(x, y, z) where (x, y, z) covers the 3D image domain, then the path of the probe is a path moving uphill in the function field f(x, y, z) to the nearest mode. This modeseeking procedure is a more straightforward version of meanshift. By starting the mode-seeking procedure from many seed points uniformly distributed within a VOI, we can detect all the SUV-maxima points that correspond to the primary locations of the true hot spots. The second step of the algorithm is to apply 3D region growing with seed points at the detected SUV-maxima points. The region growing process stops when the SUV values of the boundary points drop to 40% of the maximum SUV value at the initial seed point; this is a clinically meaningful threshold that have been utilized in the literature. We apply the mode-seeking region growing algorithm on all the VOI detected by CD, and an example segmentation result is shown in Fig. 5(c).

Fig. 5. Refining segmentation using MSRG

3.    Experiment results

Our data are acquired by SIEMENS Biograph LSO PET/CT scanners. We tested 30 cases for body volume segmentationand body section labeling (11 cases are used for training). Figure 6 shows some example results and the algorithm is robust to noise and can adapt to patients with different diseases. We applied Competition Diffusion and mode-seeking region growing algorithms to segment hot spots in all cases. Some example 3D-view segmentation results are shown in Figure 7. The segmentation process is completely automatic and the results are reproducible.

Fig. 6. Body segmentation and section labeling results

Fig. 7. Hot spot segmentation results

4.    Conclusion

We have presented a PET-CAD system, which mainly includes the following contributions: the adoption of SUV for PET data normalization, HMM for body section recognition, Context-based VOI detection with the Competition Diffusion algorithm, and clinically meaningful hot spot segmentation using mode-seeking region growing. In the future, each step of our system could be improved by utilizing CT information using PET-CT data and PET-CT registration, in order to improve system performance and provide more information to the radiologists. The recognition of abnormal vs. normal hot spots also needs more elaborate solutions.

Publication

Haiying Guan, Toshiro Kubota, Xiaolei Huang, Xiang Sean Zhou, and Matthew Turk, “Automatic Hot Spot Detection and Segmentation in Whole Body FDG-PET Images”, In Proc. of the IEEE International Conference on Image Processing (ICIP'06), pp. 85-88, 2006. (PDF)(Slides)

Patent

Xiaolei Huang, Xiang Zhou, Anna Jerebko, Arun Krishnan, Haiying Guan, Toshiro Kubota, Vaclav Potesil, “System and Method for Whole Body Landmark Detection, Segmentation and Change Quantification in Digital Images”, Patent pending, filed in Oct. 2006, U.S. Serial No. 11/542,477.