SCIENTIFIC SESSION ABSTRACT - mHealth, HIT 2.0 & the Other-ologies

Using the Microsoft Kinect for Patient Size Estimation and Radiation Dose Normalization: Proof of Concept and Initial Validation

Author:

Tessa S. Cook, MD, PhD, Hospital of the University of Pennsylvania; Gregory Couch, PhD; Timothy Couch; Woojin Kim, MD; William W. Boonn, MD

Hypothesis:

We hypothesize that the depth information provided by the Microsoft Kinect will provide an accurate estimate of whole-body volume. We assume that the density of the body estimated from this volume calculation will be close to 1000 kg/m3, which is the density of water. We also hypothesize that arm position will cause variations in the volume estimation, specifically, that volume estimates with the arms crossed over the chest will differ from volume estimates when the arms are by the sides or extended above the head but not overlapping with the body.

Introduction:

There is growing interest in being able to accurately and routinely monitor patients' exposure to ionizing radiation. Historically, whole-body effective dose has been estimated by multiplying the dose-length product (DLP) by the anatomy-specific conversion factor, or k factor, derived from tissue-specific weighting factors determined by the International Commission for Radiological Protection (ICRP). While deriving effective dose from DLP provides a straightforward, practical estimate of patients’ radiation exposure, it does not reflect patient size [1]. Using Monte Carlo simulations of CT scans and subsequent calculation of organ doses, researchers have demonstrated that DLP can underestimate dose for smaller patients, including children, and overestimate dose for larger adults [2, 3]. It is critical to understand that CTDI and the associated dose indices do not represent actual patient dose [4]. This motivated the AAPM to develop correction factors for CTDIvol based on effective patient diameter [5]. Additional work has been done to normalize for inherent differences in scanner geometries and enable comparisons of dose estimates between scanners, however, these corrections still do not account for patient factors---size, gender, body habitus [6, 7]. Recent work using Monte Carlo simulations to calculate organ doses from anthropomorphic phantoms of different sizes has more clearly illustrated how much DLP can vary with patient size [8, 9]. However, computational needs render it impractical to model every patient individually, necessitating the need for standardized phantoms.

There are many potential estimates of patient size--height, weight, body mass index (BMI), effective diameter of the anatomy being imaged. However, none of these measures effectively captures the size of the region of anatomy being imaged. To address this limitation, we present and validate a novel approach to estimating patient volume using the Microsoft Kinect, a combination RGB camera-infrared depth sensor device.

Methods:

The algorithm, developed by Radimetrics Inc., registers a depth map and computes a skeletonization of an individual who appears in front of the Kinect camera-sensor. The combination of these data is used to estimate the volume of the individual. The depth map does not display or record any identifiable features of the person, and is not saved after he/she leaves the viewing frame (Figure 1). Volume is calculated in real-time while the individual remains in front of the device. The volume estimate is continuously refined and recalculated as long as the individual remains within the camera’s field of view.

Figure 1

To validate the feasibility of using this Kinect camera-sensor for estimating patient size, we compare the calculated volume of seven non-patient volunteers and compare this value to each individual’s weight. No identifiable information about the volunteers is stored; weights and volumes are recorded anonymously. The effect of position on the volume estimation is also evaluated by estimating each volunteer’s whole-body volume with the arms by their sides, extended above the head or crossed over the chest.

Results:

Weight and volume estimates for a total of seven non-patient volunteers (five adults, two children; four males, three females) were compared. The volume estimates were obtained with each individual positioned approximately 2.5 meters in front of the Kinect camera system, standing on a scale to register their weight. To test the repeatability of the volume estimates, multiple depth images were analyzed for each individual. Repeated volume estimates were noted to be within one significant digit. Figure 2 demonstrates the depth map produced by the system for six of the volunteers – two adults and four children. In this figure, the volunteers are all standing with their arms to the sides. Figure 3 demonstrates the additional arm positions that are evaluated: arms extended upward (left) and arms crossed over the body (right).

Figure 2

Figure 3

Figure 4 compares the volume estimates computed for the three different arm configurations. Not surprisingly, there is a clear difference between the volume estimates in the two children (volunteers 2 and 3) and the remaining adults. Volume is observed to increase when the arms are extended over the head and decrease when they are crossed in front of the body. This trend is also noted in the density estimates.

Figure 4

Density is calculated by dividing each individual’s weight by his or her volume. We hypothesized that the density of each individual would be close to 1000 kg/m3, or the density of water, as the human body is primarily composed of water. The average density for the five adult volunteers is 995.2 kg/m3 with the arms by the sides, 805.1 kg/m3 with the arms extended upwards and 1136.8 kg/m3 with the arms crossed over the chest. When the arms are extended overhead, the body appears to occupy a larger volume but a higher proportion of that volume appears to be composed of air, thus resulting in a lower density estimate. Similarly, when the arms are crossed over the body, the body occupies a lower volume but appears to be composed not of air, thus increasing the estimates. Some of this variation may arise from the overall angulation of the body that can occur out of the coronal plane when the arms are positioned differently. Nevertheless, this variation with position motivates a more robust volume estimation that can account for these positional differences.

The density estimates for the two children were nearly 50% of that of the adult volunteers for each analyzed position. This was not an expected result and may be secondary to a true density difference between adults and children or the need for more precise centering in the longitudinal (head-to-toe) direction of shorter individuals in front of the Kinect camera.

Discussion:

Estimation of patient volume using the depth information produced by the Microsoft Kinect is demonstrated to be repeatable, and provides equivalent information to patient weight in terms of consistency. However, the whole-body volume estimate can provide additional information about the “thickness” of a patient, which is relevant not only to how much dose the patient receives during an imaging study but also to how much radiation is necessary to produce diagnostic-quality images. While this volume estimate is of the entire body of an individual, it can be matched to the region of the patient being imaged in order to provide a regional size estimate and ultimately more accurate size-corrected radiation dose estimates. In addition, it provides valuable feedback about the relationship between the extremities and the torso, as well as the dimensions of each, which cannot be derived from weight alone and may impact proper positioning of the patient for a complicated examination (e.g., imaging of multiple discontinuous body parts in trauma patients).

Not surprisingly, volume estimation varies with patient positioning. The effect of position on volume estimation will be explored in future work, specifically with respect to upright versus supine positioning and the variable positioning of the arms with respect to the torso. To improve the accuracy and consistency of volume estimates across the different positions, perpendicular depth maps may be necessary, obtained either with two Kinect cameras positioned at a 90-degree angle to one another, or by recording depth maps with the patient in consecutive frontal and lateral positions with respect to a single camera. Additional work is also needed to further explore the differences introduced by patient height and age, and to determine what additional modifications are necessary to accurately estimate volume in these individuals.

Conclusion:

As the need for more accurate dose monitoring increases, the best surrogates for patient size are being sought. While weight and body mass index are easily obtained, they do not always provide an accurate estimate of patient size, particularly of the region of the patient being imaged. Using the preliminary results from this novel method for estimating patient size, we can work towards more accurate means of estimating patient volume. These improved estimates of patient size can ultimately be used to correct CT dose estimates to better reflect both regional and global patient body habitus as well as patient positioning within the scanner bore. Ultimately, more accurate dose estimation improves care of our radiology patients by supplementing and improving the information that radiologists and non-radiologist physicians can use to order and protocol imaging studies appropriately.

References:

  1. C. H. McCollough, S. Leng, L. Yu, D. D. Cody, J. M. Boone, M. F. McNitt-Gray. “CT Dose Index and Patient Dose: They Are Not the Same Thing”. Radiology 259: 311-316, 2011.
  2. D. J. Brenner, C. H. McCollough, C. G. Orton. “Is it time to retire the CTDI for CT quality assurance and dose optimization?” Med Phys. 33: 1189-91, 2006.
  3. L. M. Hurwitz, T. T. Yoshizumi, P. C. Goodman, et al. “Effective Dose Determination Using an Anthropomorphic Phantom and Metal Oxide Semiconductor Field Effect Transistor Technology for Clinical Adult Body Multidetector Array Computed Tomography Protocols.” J Comput Assist Tomogr. 31(4): 544-549, 2007.
  4. C. H. McCollough, S. Leng, L. Yu, et al. “CT Dose Index and Patient Dose: They Are Not the Same Thing.” Radiology 259(2): 311-316, 2011.
  5. American Association of Physicists in Medicine (2011). “Size-Specific Dose Estimates (SSDE) in Pediatric and Adult Body CT Examinations (Report #204)”. College Park, MD. (Implemented by TG-204)
  6. A. C. Turner, M. Zankl, J. J. DeMarco, et al. “The feasibility of a scanner-independent technique to estimate organ dose from MDCT scans: Using CTDIvol to account for differences between scanners.” Med Phys. 37(4): 1816-1825, 2010.
  7. W. Huda, F. A. Mettler. “Volume CT Dose Index and Dose-Length Product Displayed...What Good Are They?” Radiology 258: 236-242, 2011.
  8. X. Li, W. P. Segars, E. Samei, et al. “Patient-specific dose estimation for pediatric chest CT”. Med Phys. 35: 5821-28, 2008.
  9. X. Li, E. Samei, W. P. Segars, et al. “Patient-specific radiation dose and cancer risk estimation in CT: Part II. Application to patients.” Med Phys. 38: 408-419, 2011.

Keywords:

CT dose monitoring
Size-corrected dose estimation
Patient-specific dose estimates