Estimating normal metabolic activity for disease quantification via PET/CT images

2.1. DQ pipeline

A schematic representation of the proposed DQ pipeline via PET/CT images is shown in Figure 1. The pipeline consists of three key components: SUV standardization, object segmentation, and calculation of disease quantity within the object. Metabolic and anatomical information required in the DQ process is separately derived from the matched pair of PET and CT images.

To deal with the problem of inconsistent numeric pixel meaning in PET images, SUV images are usually generated by normalizing the PET values with injected radiotracer dose and patient body weight. Our previous work [11] further proposed a non-linear method of standardizing the overall scale of (near) whole-body PET images to alleviate the influence on image acquisition from other unquantifiable factors.

As most segmentation algorithms are explored in specific body regions with consideration for both accuracy and time-efficiency, body region recognition should be first conducted on the (near) whole-body CT images to extract the slice stack of interest [12]. Then, masks of the target objects can be obtained from the CT images by various segmentation algorithms, such as one of our previously proposed methods which combines fuzzy anatomy modeling for recognition and deep learning networks for delineation [13].

2.2. Calculating the total lesion burden

Compared with commonly used quantitative measures such as SUV_max, SUV_mean, and SUV_peak, which show stable results under the repeatability problem of PET image acquisition and challenges in lesion segmentation [6], total lesion burden (TLB) may show better comparability and overall clinical meaning in practice. TLB(O) for an object O is calculated by estimating the background normal metabolic activity (TMA_n) and subtracting it from the total metabolic activity (TMA) within the target object in the standardized SUV image, as shown in Equation (1).

For any image I and for any object O, we define total metabolic activity TMA(O) of O as in Equation (2).

where I_S stands for the standardized SUV image, v stands for each voxel in the image, and FM_O stands for the segmentation map of the target object O which can be either a fuzzy map (0 to 1, representing the probability or membership of the voxel belonging to the object) or a binary map (0 or 1, representing background and foreground separately).

TMA_n is calculated with SUV models generated from a modeling dataset obtained from disease-free subjects with no obvious lesion having abnormal radiotracer uptake in the PET images. As the patterns of radiotracer uptake may be different among organs under consideration, SUV model or normality distribution, denoted by ND_O(s) for SUV value as a random variable s, is generated for each of the organs separately. TMA_n is explored in hard and fuzzy ways, correspondingly denoted by ${\text{TMA}}_{\text{n}}^{\text{h}}(\text{O})$ and ${\text{TMA}}_{\text{n}}^{\text{f}}(\text{O})$.

3.1. Experiments

The proposed DQ method is evaluated on clinical and phantom datasets separately. The retrospective study was conducted following approval from the Institutional Review Board at the Hospital of the University of Pennsylvania along with a Health Insurance Portability and Accountability Act waiver. The thoracic regions of 34 ¹⁸F-fluorodeoxyglucose (FDG)-PET/CT scans from normal subjects are utilized in the evaluation experiments. Masks of 4 objects are manually delineated on the CT images, including left lung (LLg), right lung (RLg), thoracic esophagus (TEs), and thoracic skeleton (TSk), as in the example shown in Figure 2(a) where the RLg mask is overlaid on the PET image. DQ on lesions is evaluated on 20 samples of the SNM Germanium Phantom datasets, where the phantom was filled with ⁶⁸Ge (271-day half-life) in an epoxy matrix and the data were acquired using a GE Discovery STE-16 PET/CT scanner (with 16-row MDCT) [14]. Lesions are represented by spheres of 6 different sizes, respectively with diameters of 10 mm, 13 mm, 17 mm, 22 mm, 28 mm, and 37 mm. Masks of normal background and lesions are generated by thresholds as shown in Figure 2(b). When conducting DQ on a single lesion, a box extended by 4 mm beyond the bounding box of the binary lesion is generated for each lesion, representing a local region of interest containing both normal background and lesion, as shown in Figure 2(c). In the stage of model building, 20 out of 34 samples and 10 out of 20 samples are selected as the modeling datasets, where the remaining 14 and 10 samples are taken as the testing datasets. Histograms of the 4 thoracic organs and background normal tissue of phantom are created for generating SUV models as in (4).

3.2. Results and discussions

The results are summarized in Table 1. The ratio of absolute TLB to TMA is utilized as the evaluation metric, which alleviates the individual influence of both volume size and uptake level of the sample. We also compare the ratios of the results derived from hard and fuzzy strategies for estimating TMA_n.

The fuzzy strategy shows better performance in both actual datasets and phantoms. For normal organs in the clinical PET image datasets, as well as the background tissue in the phantom PET image datasets, the ratios of absolute TLB to TMA are close to 0 representing disease-free regions. The ratios estimated by the fuzzy strategy show overall lower mean values than the hard strategy, except in the case of phantom normal background region while the fuzzy strategy reaches a standard deviation value lower than 0.001. For phantom lesions, TLB values are calculated within the extended boxes and TMA values are determined in the lesion region only. The ratios are close to 1, representing that the contribution to TLB within the boxes is minimal from the background normal tissue and mainly from the lesion region. Moreover, for the lesions with larger diameters, the ratio also reaches closer to 1, showing degrading accuracy with decreased ability to account for partial volume effect in smaller lesions.