PAPER

J Forensic Sci, 2014 doi: 10.1111/1556-4029.12531 Available online at: onlinelibrary.wiley.com

ANTHROPOLOGY Sharon M. Derrick,1 Ph.D.; Michelle H. Raxter,1 Ph.D.; John A. Hipp,2 Ph.D.; Priya Goel,2 M.S.; Elaine F. Chan,2 Ph.D.; Jennifer C. Love,1 Ph.D.; Jason M. Wiersema,1 Ph.D.; and N. Shastry Akella,2 Ph.D.

Development of a Computer-Assisted Forensic Radiographic Identification Method Using the Lateral Cervical and Lumbar Spine*

ABSTRACT: Medical examiners and coroners (ME/C) in the United States hold statutory responsibility to identify deceased individuals who

fall under their jurisdiction. The computer-assisted decedent identification (CADI) project was designed to modify software used in diagnosis and treatment of spinal injuries into a mathematically validated tool for ME/C identification of fleshed decedents. CADI software analyzes the shapes of targeted vertebral bodies imaged in an array of standard radiographs and quantifies the likelihood that any two of the radiographs contain matching vertebral bodies. Six validation tests measured the repeatability, reliability, and sensitivity of the method, and the effects of age, sex, and number of radiographs in array composition. CADI returned a 92–100% success rate in identifying the true matching pair of vertebrae within arrays of five to 30 radiographs. Further development of CADI is expected to produce a novel identification method for use in ME/C offices that is reliable, timely, and cost-effective.

KEYWORDS: forensic science, forensic anthropology, radiographic identification, cervical, lumbar, vertebrae

Medical examiners and coroners (ME/C) in the United States hold statutory responsibility for the identification of decedents that fall under their medicolegal jurisdiction. Due to heavy ME/ C caseloads, pervasive budget constraints, concern for delays experienced by the decedent’s family, and storage issues, the preferred identification methods are those that provide not only reliable, but timely and cost-effective results. A quick turnaround time is especially important when the decedent presents to the ME/C with a tentative name, but the identity must be scientifically confirmed due to death circumstances, such as multiple fatalities where there is potential for incorrect assignment of identity, disfiguring trauma or decomposition, or the requirements for certain homicide cases. Delays in identification of these decedents may be confusing and frustrating to the family because it may be difficult for family members to understand why the body cannot be released immediately postautopsy for burial or cremation. Further, the decedent must be stored for a period of time under morgue refrigeration, which is not an ideal

1 Harris County Institute of Forensic Sciences, 1885 Old Spanish Trail, Houston, TX 77054. 2 Medical Metrics, Inc., 2121 Sage Road #300, Houston TX 77056. *Presented in part at the 63rd Annual Meeting of the American Academy of Forensic Sciences, February 21–26, 2011, in Chicago, IL; at the 64th Annual Meeting of the American Academy of Forensic Sciences, February 20–25, 2013, in Atlanta, GA; and at the 65th Annual Meeting of the American Academy of Forensic Sciences, February 18–23, 2013, in Washington, DC. Received 29 May 2013; and in revised form 6 Sept. 2013; accepted 14 Sept. 2013.

© 2014 American Academy of Forensic Sciences

situation for the ME/C office, the family, or the receiving funeral home. When determining identity, the ME/C should be cognizant that evidentiary standards for forensic evidence, including those used for decedent identifications, have changed in the post-Daubert era (1–4). Judicial inquiry regarding the admissibility of expert testimony places more emphasis on the reliability of the scientific methods used by practitioners (5). Developers and practitioners of forensic identification methods must ensure that the results and interpretations are evidence based and potentially admissible in court given the admissibility criteria for expert testimony established in Supreme Court cases (1–4). At the state level, researchers have noted that 91% of surveyed state court judges indicated quantified error rates to be useful in their assessment of the quality of presented scientific evidence (6). Additionally, an influential report released by the National Academy of Sciences in 2009 (7) emphasizes the importance to the forensic sciences of a firm foundation in the scientific method. The most common forensic identification methods currently used in the ME/C setting consist of fingerprinting, dental charting and dental radiograph comparison, skeletal radiographic comparison and evidence of medical intervention, and DNA identification-based profile comparison. Each of these methods is based in biometric analysis and each has advantages and disadvantages. Fingerprinting, although timely and inexpensive, may not be possible when the body is thermally modified, traumatized, or in an advanced stage of decomposition as it is relatively common for decedent fingerprints to be obliterated. Furthermore, antemortem fingerprint records may not be available for comparison. Additionally, when computerized fingerprint comparison 1

2

JOURNAL OF FORENSIC SCIENCES

through the automated fingerprint identification system (AFIS) is not successful, latent examination may be required. Recent research suggests that examiner fingerprint comparison may be adversely affected by confirmation bias, which may lead to evidentiary challenges (8–11). Dental identification is most reliable when the decedent has antemortem dental records available for comparison with postmortem examination (12). Lack of antemortem dental care and/ or archived records as well as antemortem or postmortem tooth trauma or loss may preclude adequate comparison of dental characteristics. When DNA profile comparison is deemed necessary, a direct or family reference sample may not be available, and remains that are subject to thermal or chemical degradation may not contain adequate DNA for analysis. Forensic anthropologists are frequently called upon by the ME/C to use their expertise in skeletal anatomy and morphology to compare antemortem and postmortem radiographs for the purposes of identification. The forensic anthropologist typically undertakes identification analysis by qualitative comparison of skeletal elements imaged in antemortem and postmortem radiographs. An anthropologist experienced in this method can be highly accurate (13,14). However, the validity of the anthropologist’s conclusions may be affected by the individual’s training and experience level. Further, the actual individuality of the element shapes, trabecular patterns, and other osseous features used in identification has not been quantified, nor has the effect that a small deviation in radiograph positioning may have on the shape as seen by the analyst. As a result, an expected error rate, one of the components of the Daubert test, cannot be adequately defined. Published literature on identification by anthropologic radiograph comparison indicates that although there is general peer acceptance of the methodology, there is a lack of confidence that identifications made using traditional radiograph comparison will be admissible in court in light of the recent revisions of admissibility standards (4,15). In response, researchers have begun to explore avenues to strengthen the statistical support for radiograph comparison in various regions of the skeleton, in disarticulated remains as well as fleshed remains (14–23). For example, Stephan et al. (16) quantified a 55% gap in accuracy between trained and untrained examiners in their comparison of antemortem chest radiographs with postmortem radiographs of disarticulated skeletal elements fixed in anatomic position. Harris County Institute of Forensic Sciences (HCIFS) located in Houston, Texas, serves a large urban-based county of 1703 square miles with an estimated four million residents (24). In 2011, HCIFS performed autopsies or external examinations on 3818 cases. Of the 3818 cases, 28% (1101) required forensic identification, with 576 of these decedents tentatively identified at check-in. The tentatively identified cases included fleshed decomposing remains, charred remains, and commingled assemblages from motor vehicle and airplane crashes, homeless individuals without documentation of identity, and presumptive homicide cases. HCIFS standard operating procedure directs a sequential process for scientific identification: fingerprint comparison by outside agencies; radiograph comparison by in-house anthropologists; dental comparison by the consulting odontologist; and DNA comparison by the in-house laboratory. The process of identification is both time and cost dependent. In 2011, 943 decedents were identified using fingerprints from Texas driver licenses or criminal records at no billed cost to HCIFS. The turnaround time was a few hours to c. 2 days. Twenty-five

decedents were identified by in-house anthropologists within hours at no additional cost to HCIFS. Fifty-seven were identified by odontologic examination at a total consultant cost of $24,000 and a turnaround time of 2–3 days. Finally, 76 decedents were identified through DNA profile comparison at an estimated internal cost of $4560, and with a routine time delay of 15–60 days from submission of the sample. ME/C offices without an on-site DNA laboratory are likely to experience a longer turnaround interval for DNA results. External DNA analysis, previously expensive but currently offered at no charge through the President’s DNA Initiative (25), has a turnaround time historically measured in months. HCIFS data indicate that the implementation of a validated and statistically quantifiable method of identification that reduces the need for subjective radiograph comparison by experts, costly odontologic examination, and time-consuming DNA profile analysis would be an important resource for large ME/C offices. The novel method presented here, computer-assisted decedent identification (CADI), is an automated forensic personal identification tool that is nondestructive and does not require the handling of biohazardous materials beyond standard decedent manipulation. CADI was developed for routine use in the ME/C office through collaboration among Harris County Institute of Forensic Sciences forensic anthropologists and biomedical engineers, software engineers, and software designers affiliated with Medical Metrics, Inc., (MMI). The goal of the study was to develop the CADI software, subsequently test its accuracy in analyzing the shapes of targeted skeletal elements imaged in an array of standard radiographs, and assess its ability to quantify the likelihood that any two of the radiographs contain elements that belong to the same individual. The foundation of CADI is a forensic version of Quantitative Motion Analysisâ software (QMAâ) developed by MMI, Inc., Houston, Texas, USA. MMI is an independent imaging core laboratory specializing in biomedical image data management, radiographic analysis, and consulting services. The company was founded using technology, including QMAâ, developed at Baylor College of Medicine, Houston, Texas. The original QMAâ allows for computer-assisted matching of specific objects that can be seen in multiple radiographic images. QMAâ has been validated in multiple studies (26–29) and used in over 100 peerreviewed studies of spinal biomechanics and spinal treatments (reference list available on request). The software tracks a specific object between radiographic images and has most commonly been used in lateral views to measure motion between vertebrae as well as changes in spatial relationships between vertebrae over time. The QMAâ ability to track objects is a fundamental feature of CADI that facilitates determination of whether a specific anatomical feature is a match, as defined below, in radiographs taken at different time periods (e.g., antemortem and postmortem images). Specifically, the software is used to position all images in an array so that the feature of interest is in the same orientation and size (as best as can be achieved) in all images in the array.

Methods CADI Software Development Computer-assisted decedent identification method development began with an assessment of de-identified lateral cervical and lumbar (N = 11, C3-5; N = 30, L1-5) spine radiographs previously archived at MMI from multiple clinical trials of spine

DERRICK ET AL.

treatments. The initial goal was to determine the optimal filters and match algorithms required to perform the automated comparison of skeletal element shape. The sample images were randomly selected based on the subject’s age, sex, and on the availability of two radiographs of the same vertebral region taken 2–4 years apart. Prior to testing, the compiled radiographs were reviewed for opacity artifacts, out of plane artifacts, excessive noise, poor contrast, or obstruction that might eliminate the radiograph from use in an actual decedent identification case. Radiographs of subjects with severe osteophytic growth between scans or with a visible surgical implant were also excluded. Custom codes were written to provide an analyst-friendly interface, perform the image processing, and generate similarity metrics using MATLAB (Mathworks, Natick, MA). The images were spatially stabilized and scaled by CADI so that the endplates of each vertebra were the same approximate size. Although the size of an element may differ markedly between individuals, sheer size of the element is not used for matching purposes in this method because antemortem and postmortem radiographs will rarely be of the same magnification, and the magnification for each radiograph is not known. Therefore, CADI adjusts all images to minimize differences in magnification such that the size of the element is similar in all images in the array. Computer-assisted decedent identification extracted and preprocessed the images to adjust intensity, correct for streak artifacts, reduce noise, and adjust the image resolution. This step minimizes the effect of variability in image quality to avoid complications in the assessment of similarities and differences. A polygon region of interest around the vertebral body was manually defined by the analyst to mask out everything but the target vertebral body (Figs 1 and 2). Each test array was configured to include 12 radiographs: one Identification pair (ID pair) and 10 comparison images (N = 80 cervical and 90 lumbar radiograph arrays). The ID pair was comprised of an early time point radiograph (simulating antemortem) and a radiograph taken 2 to 4 years later (simulating postmortem) from the same individual. To construct an array, the ID pair was combined with radiographs from 10 different individuals (one radiograph per individual) and used to represent a pool of other possible antemortem radiographs from the same sex and age cohort. Thus, each test array consisted of one correct match pair and 10 potential incorrect matches (Fig. 3). Five image-processing filters (None, Histogram Equalization, Adaptive Histogram Equalization, North filter, Kalman Stack filter) and five image-match algorithms (Dice, Jaccard, Structural Similarity, Mutual Information, Matrix Correlation) were tested for their ability to delineate the shape characteristics of the vertebral body and generate the similarity metric, or the match score, for a pair of stabilized radiographs (30–38). For each array, match scores between the postmortem image of the ID pair, the antemortem image of the ID pair, and all comparison radiographs in the array were calculated. Each combination of imageprocessing filter and match score algorithm was tested (Fig. 4). Match scores were statistically analyzed using Stata (StataCorp, College Station, TX). For each combination of a preprocessing filter and match score algorithm, the percentage of true positives and false positives was computed. The percent correctly matched was calculated using receiver–operator curve (ROC) analysis. In addition, an effect size was calculated (average match score of true positives minus average match score of false positives, divided by the standard deviation for all match scores). The optimal combi-

.

DEVELOPMENT OF CADI

3

FIG. 1––Fundamental steps of the CADI process for radiographic array comparison and match score calculation.

nation of image-processing filter and match score algorithm was determined from the pooled comparisons. CADI generated >90% correctly classified results with any combination of image-processing filter and match score algorithm. For the cervical region, the optimal filter and algorithm combination were a histogram equalization filter (improves contrast when the pixel value distribution is similar across the image) followed by the Jaccard algorithm similarity coefficient (statistical measure of the similarity between sample sets). For the lumbar region, the optimal filter and match score algorithm were adaptive histogram equalization filter (improves contrast when the pixel value distribution is not uniform across the image) followed by the Jaccard algorithm similarity coefficient. These combinations of filter and match score algorithm produced >93% accuracy in correctly classifying ID pairs (98% for cervical, 93.61% for lumbar). These respective filters and match score algorithm combinations were used for subsequent testing of the software. CADI Testing Once the optimal filters and match score algorithms were identified, the project moved forward with further coding revision of QMAâ into the CADI software prototype and validation of CADI for the cervical and lumbar anatomical regions using radiographs from the MMI repository and radiographs of the same regions from the second National Health and Nutrition Examination Survey (NHANES II) (39). Multiple age and sexspecific arrays containing 5–30 images and at least one ID pair

4

JOURNAL OF FORENSIC SCIENCES

spatially stabilized and preprocessed by the software as described above for all images in the test array (Fig. 3). The antemortem and postmortem images were masked for each test to isolate only the feature of interest (Fig. 2). In order to validate the CADI prototype, multiple test scenarios were executed in validation tests 1–6. The tests were constructed to evaluate the following: (i) whether the performance of the software varies with age and sex (Test 1 and Test 3), (ii) what effects result from varying the number of radiographs included in a test array (Test 2), (iii) the repeatability and reliability of the tracking and masking processes (Test 4), (iv) the effects of varying the pool of comparison radiographs in a test array (Test 5) and (v) the sensitivity of the method when multiple correct matches are present in a test array (Test 6). The validation tests were followed by a brief test of CADI accuracy when compared to the accuracy of six forensic anthropologists assessing the same radiograph arrays.

Results Validation Tests 1–6

FIG. 2––A polygon region of interest is drawn around the vertebral body to delineate the target for CADI analysis. The top image is a simulated postmortem radiograph and the bottom image is a simulated antemortem radiograph. Both images are from the same individual.

FIG. 3––A sample test array of twelve preprocessed radiographs of the L1 vertebral body in which two of the radiographs represent the ID pair.

were assembled. The correctly matched radiographs (the ID pairs) used to represent antemortem and postmortem pairs were selected from the MMI repository with the requirement that the radiographs were taken at least 2 years apart. The unmatched cases were randomly selected for each test array based on age and sex of the correctly matched radiographs. Arrays were constructed for males and females in 5-year-age cohorts within an age range of 40–69 years. A skeletal element of interest was

In Test 1, age and sex-specific arrays were created using radiographs compiled from two separate clinical studies in order to analyze match scores separately by sex, age, and then sex and age combined. Within each age- and sex-specific group, the influence of individual vertebral level on the correct classification of radiographs was also tested. In total, individual cervical vertebrae were correctly classified in 98.4% of males and 98.9% of females. Individual lumbar vertebrae were correctly classified in 94.1% of males and 93.2% of females (Tables 1–3). In Test 2, the influence of array size on correct classification was investigated using the male, 40–44 years, C3 level Test Array from Test 1. The C3 level was chosen for Test 2 because the position of C3 within the spine increased the availability of suitable radiographs in the MMI radiograph archives to populate the test arrays. Test 2 included five ID pairs and a varying number of radiographs from different individuals (5–30 radiographs in increments of 5). All radiographs (100%) were correctly classified in this test regardless of the array size (Table 4). The importance of nonage-specific test arrays on correct classification was determined by Test 3. Two radiographs from different male individuals were randomly selected from each of the six age groups. The 12 radiographs formed a base array that spanned the entire age range from 40–69 years. One ID pair from each age group was added to the base array independently, forming six final test arrays. All radiographs (100%) were correctly classified in Test 3 (Table 5). Test 4 examined the repeatability and reliability of the tracking process using QMAâ and the masking performed by the analyst. Radiographs for the male, 40–44 years, C3 level test array from Test 1 was preprocessed (stabilization and masking) and re-tested on two separate occasions. All radiographs (100%) were correctly classified in both events of Test 4. Test 5 found that varying the pool of comparison images did not affect correct classification of radiographs. All of the radiographs (100%) were correctly classified using two test arrays with different individual radiographs paired to the same ID pair. Test 6 determined whether multiple correct matches could be identified in a single test array. An age and sex-specific test array was created with five ID pairs and five incorrect match radiographs. Thus, each of the five ID pairs was tested against nine incorrect match radiographs. All five of the ID pairs

DERRICK ET AL.

.

DEVELOPMENT OF CADI

5

FIG. 4––Testing matrix for each combination of image processing filter and match score algorithm (30–38).

(100%) were correctly matched. In summary, the cervical vertebrae results ranged from 95% to 100% of the ID pairs correctly matched. The lumbar results ranged from 92% to 99% of the ID pairs correctly matched. CADI Results Compared with Forensic Anthropologist Findings Following compilation of the CADI cervical and lumbar vertebrae results, a brief comparison of CADI accuracy with traditional radiograph comparison was performed. Six forensic anthropologists (three with minimal experience and three highly experienced in radiograph comparison for identification) assessed 27 cervical test arrays that had been successfully analyzed by CADI (100% correct). Test arrays with no matching radiographs were included in the exercise. The anthropologists were allowed to use all morphological features observed in the lateral radiographs of the cervical spine. The anthropologists’ results ranged from 88.9% to 100% of the ID pairs correctly identified, slightly lower than the results generated by CADI for the bodies only. The performance of CADI in this test indicates that the prototype software is as accurate for lateral cervical radiographs as traditional radiograph comparison performed by forensic anthropologists. Additionally, CADI produced a calculated match score to support each comparison result.

TABLE 1––Test arrays created for age- and sex-specific groups and the vertebral levels analyzed. The ID pair and comparison radiographs were compiled from two separate clinical trial projects. Age Range (years) 40–44 45–49 50–54 55–59 60–64 65–69 40–44 45–49 50–54 55–59 60–64 65–69

Sex Male Male Male Male Male Male Female Female Female Female Female Female

Cervical Vertebral Level C3, C4, C3, C4, C3, C4, C3, C4, C3, C4, C3, C4 C3, C4, C3, C4, C3, C4, C3, C4, C3, C4, C3, C4

C5 C5 C5 C5 C5 C5 C5 C5 C5 C5

Lumbar Vertebral Level L2, L1, L1, L1, L1, L2, L1, L2, L1, L1, L1, L1,

L3, L2, L2, L2, L2, L3, L2, L3, L2, L2, L2, L2,

L4 L3, L3, L3, L3, L4 L3, L4, L3, L3, L3, L3,

L4 L4 L4, L5 L4, L5 L4 L5 L4, L5 L4 L4 L4, L5

Discussion The CADI project was designed to test the hypothesis that biometric-based software currently used in the diagnosis and treatment of spinal injuries could be modified and validated as an objective forensic tool for identification. Biometric verification of identity is in widespread use throughout various scientific and commercial industries. Analysis and quantification of features, such as hand shape and finger length, and the network of capillary vessels viewed by retinal scan, are used to authenticate an individual’s identity for access to secured information and physical locations (40). Picture archiving and communication system (PACS), routinely utilized in major hospital networks and in the individual clinical setting (41,42), is a biometric-based system that conducts automated shape matching of edgeenhanced radiographs to ensure that the radiograph is archived in the correct patient file. Morishita et al. (42), in their work with chest radiograph matching for patient recognition, described the method as a form of “biological fingerprint”. More recently, work by van der Meer et al. (12) used digital subtraction radiography (ImageTool v3.0 with UT-ID plug-in, University of Texas Health Science Center, San Antonio, Texas) to test a potential means of automated dental comparison for identification purposes. Questions of reliability/replicability may arise when a method relies on a nonobjective foundation. Christensen and Crowder (43) point out that experience level is important in the interpretation of analytical results, but experience should not dictate the validity of the scientific methods used to obtain the results. Rogers and Allard (44) predicate that “A mathematical means of arriving at a positive identification ensures replicability, makes criteria explicit, and provides a method that can be debated and discussed”. CADI has been statistically validated to correctly match individual cervical and lumbar vertebral bodies in lateral radiographic views taken of the same individual at different times. Based on the results of the current experiments, age and sex-matched arrays are not essential to using a method such as CADI for decedent identification. However, in actual ME/C identification casework, it is likely that the confidence in the results will be improved if age and sex-matched arrays are used. The CADI system can maintain a large pool of radiographs such that age and sex-matched arrays can be easily assembled as needed.

6

JOURNAL OF FORENSIC SCIENCES

TABLE 2––Influence of sex, cervical vertebral level, or age and sex on the percentage of correctly classified radiographs utilizing the histogram equalization filter and the Jaccard similarity match score.

Age Range Overall All age ranges combined Sex-specific All age ranges combined All age ranges combined Vertebral level-specific All age ranges combined All age ranges combined All age ranges combined Age- and sex-specific 40–44 years 45–49 years 50–54 years 55–59 years 60–64 years 65–69 years 40–44 years 45–49 years 50–54 years 55–59 years 60–64 years 65–69 years

Sex

Cervical Vertebral Level

No. of Vertebral Levelspecific Arrays

No. of ID Pairs

% Correctly Classified

Male and Female combined

C3, C4, C5 combined

83

23 to 31

98.69

Male Female

C3, C4, C5 combined C3, C4, C5 combined

48 35

14 to 18 9 to 13

98.44 98.86

Male and Female combined Male and Female combined Male and Female combined

C3 C4 C5

31 29 23

31 29 23

98.83 99.37 98.02

Male Male Male Male Male Male Female Female Female Female Female Female

C3, C3, C3, C3, C3, C3, C3, C3, C3, C3, C3, C3,

C4, C5 combined C4, C5 combined C4, C5 combined C4, C5 combined C4, C5 combined C4 combined C4, C5 combined C4, C5 combined C4, C5 combined C4, C5 combined C4 combined C4 combined

11 9 9 9 8 2 7 9 6 9 2 2

3 to 5 3 3 3 2 to 3 1 1 to 3 3 2 3 1 1

99.17 100 100 94.95 100 100 98.70 100 100 97.98 100 100

TABLE 3––Influence of sex, lumbar vertebral level, or age and sex on the percentage of correctly classified radiographs using the adaptive histogram equalization filter and the Jaccard similarity match score.

Sex

Lumbar Vertebral Level

No. of Vertebral Levelspecific Arrays

No. of ID Pairs

% Correctly Classified

Male and Female combined

L1, L2, L3, L4, L5 combined

192

46

93.61

Male Female

L1, L2, L3, L4, L5 combined L1, L2, L3, L4, L5 combined

109 83

27 19

94.08 93.21

37 45 45 45 20

37 45 45 45 20

93.61 93.94 92.93 94.14 95.45

9 15 24 30 25 6 8 8 10 12 20 25

3 5 6 6 5 2 2 2 2 3 5 5

98.99 95.15 95.08 93.64 93.09 95.45 92.05 96.59 96.36 93.94 93.18 92.73

Age Range Overall All age ranges combined Sex-specific All age ranges combined All age ranges combined Vertebral level-specific All age ranges combined All age ranges combined All age ranges combined All age ranges combined All age ranges combined Age- and sex-specific 40–44 years 45–49 years 50–54 years 55–59 years 60–64 years 65–69 years 40–44 years 45–49 years 50–54 years 55–59 years 60–64 years 65–69 years

Male Male Male Male Male

and and and and and

Male Male Male Male Male Male Female Female Female Female Female Female

Female Female Female Female Female

combined combined combined combined combined

L1 L2 L3 L4 L5 L1, L1, L1, L1, L1, L1, L1, L1, L1, L1, L1, L1,

L2, L2, L2, L2, L2, L2, L2, L2, L2, L2, L2, L2,

L3, L3, L3, L3, L3, L3, L3, L3, L3, L3, L3, L3,

The vertebral body was used in the current study, and the arrays consisted of single vertebral bodies from the same anatomical level (i.e., all C3 or all C4). The probability of selecting the correctly matched vertebra from a pool of 11 images of vertebral bodies is 1 in 11. If the test is performed for multiple features in a radiograph (e.g., vertebral bodies of C3, C4, C5 and spinous processes of C3, C4, C5), and the correct match is found in 6 arrays constructed from a single ID pair, then the probability can be calculated as 1/11*1/11*1/11*1/11*1/11*1/ 11 = 1 in 1,771,561—although correction for repeated measures is desirable if the same array of unmatched spines is used each

L4, L4, L4, L4, L4, L4, L4, L4, L4, L4, L4, L4,

L5 L5 L5 L5 L5 L5 L5 L5 L5 L5 L5 L5

combined combined combined combined combined combined combined combined combined combined combined combined

time. This is analogous to the current practice of identifying multiple unique features when comparing radiographs for victim identification. The optimal method for interpreting match scores for multiple features within radiographs has not been established. The process of collecting and combining results for multiple anatomic features within radiographs is currently being explored. Interpretation of the range of variability seen in the raw match scores for target anatomic features is planned during further testing, revision, and validation of the prototype CADI software. Shamir et al. (45), in their radiographic identification study of the knee, noted that the recognition accuracy of the knee joint

DERRICK ET AL. TABLE 4––Influence of array size on the percentage of correctly classified radiographs using the histogram equalization filter and the Jaccard similarity match score. Age Range (Years) 40–44 40–44 40–44 40–44 40–44 40–44

Sex

No. of Comparison Images

No. of ID Pairs

% Correctly Classified

Male Male Male Male Male Male

C3 C3 C3 C3 C3 C3

5 10 15 20 25 30

5 5 5 5 5 5

100 100 100 100 100 100

TABLE 5––Influence of nonage-specific arrays on the percentage of correctly classified radiographs using the histogram equalization filter and the Jaccard similarity match score.

40–44 45–49 50–54 55–59 60–64 65–69 All age ranges combined

DEVELOPMENT OF CADI

7

the ME/C office when antemortem radiographs are available for a decedent with a tentative identity. Acknowledgments

Vertebral Level

Age Range of ID Pair (Years)*

.

Sex

Vertebral Level

No. of Comparison Images

No. of ID Pairs

% Correctly Classified

Male Male Male Male Male Male Male

C3 C3 C3 C3 C3 C3 C3

12 12 12 12 12 12 72

1 1 1 1 1 1 6

100 100 100 100 100 100 100

*Age range of the correct match ID pair of radiographs. Incorrect match radiographs spanned the entire age range from 40 to 69 years.

was statistically higher than random but became less accurate as the number of individuals in the dataset increased. These results support the inclusion of more than one anatomic region in a radiograph comparison analysis. Each new anatomical region will be individually tested to determine the best match score and the threshold levels of the match score that optimally differentiates between correct and incorrect matches. Successful validation of CADI on additional regions of the body (e.g., pelvis, hand, etc.) will increase the range of antemortem radiographs that can be utilized in practice. The magnitude of increased statistical strength obtained by combination of more than one anatomical region per decedent during an identification test is yet to be determined. Currently, the CADI system is not software that can be distributed and used outside of a central image analysis laboratory. The process of tracking anatomic features requires analyst training and certification. It is possible to provide an external training and certification program, although any potential bias that a site might exert on the results can be avoided by having the analysis performed at an independent facility. Conclusion Computer-assisted decedent identification is nondestructive and does not require the handling of biohazardous materials beyond routine decedent manipulation. The method is equipped with an analyst-friendly interface and has the potential to become a cost-effective commercial product for use by ME/C offices. Analyst subjectivity has been greatly reduced if not removed. Results generated by CADI are replicable and supported by a calculated statistical error rate. The method is efficient, timely, and may be a helpful solution to identification in

We thank Dr. Deborrah C. Pinto for her support and editing skills during preparation of the manuscript. We also thank Ms. Susan Donaldson for her graphic design assistance. We wish to acknowledge Dr. Luis A. Sanchez, Chief Medical Examiner of Harris County, for his encouragement during our research and method development. Funding provided by the National Institute of Justice (NIJ Award 2010-DN-BX-K194). References 1. Supreme Court of the United States Syllabus. DAUBERT et ux., individually and as guardians ad litem for DAUBERT, et al. v. MERRELL DOW PHARMACEUTICALS, INC. (1993); http://www.law.cornell.edu/ supct/html/92-102.ZS.html (accessed August 22, 2013). 2. Grivas CR, Komar DA. Kumho, Daubert, and the nature of scientific inquiry: implications for forensic anthropology. J Forensic Sci 2008;53 (4):771–6. 3. Keierleber JA, Bohan TL. Ten years after Daubert: the status of the states. J Forensic Sci 2005;50(5):1154–63. 4. Christensen AM. The impact of Daubert: implications for testimony and research in forensic anthropology (and the use of frontal sinuses in personal identification). J Forensic Sci 2004;49(3):427–30. 5. Moreno JA. Einstein on the bench. Ohio State Law J 2003;64:531– 85. 6. Gatowski SI, Dobbin SA, Richardson JT, Ginsburg GP, Merlino ML, Dahir V. Asking the gatekeepers: a national survey of judges on judging expert evidence in a post-Daubert world. Law Hum Behav 2001;25 (5):433–58. 7. Committee on Identifying the Needs of the Forensic Science Community. Strengthening forensic science in the United States: a path forward. Washington, DC: The National Academies Press, 2009. 8. Spinney L. The fine print. Nature 2010;464:344–6. 9. Neufeld P, Scheck B. Making forensic science more scientific. Nature 2010;464:351. 10. Dror IE, Charlton D, Peron A. Contextual information renders experts vulnerable to making erroneous identifications. Forensic Sci Int 2006;156 (1):74–8. 11. Krane DE, Ford S, Gilder JR, Inman K, Jamieson A, Koppl R, et al. Sequential unmasking: a means of minimizing observer effects in forensic DNA interpretation. J Forensic Sci 2008;53(4):1006–7. 12. van der Meer DT, Brumit PC, Schrader BA, Dove SB, Senn DR. Root morphology and anatomical patterns in forensic dental identification: a comparison of computer-aided identification with traditional forensic dental identification. J Forensic Sci 2010;55(6):1499–503. 13. Mundorf AZ, Vidoli G, Melinek J. Anthropological and radiographic comparison of vertebrae for identification of decomposed human remains. J Forensic Sci 2006;51(5):1002–4. 14. Kuehn CM, Taylor KM, Mann FA, Wilson AJ, Harruff AC. Validation of chest x-ray comparisons for unknown decedent identification. J Forensic Sci 2002;47(4):725–9. 15. Christensen AM. Testing the reliability of frontal sinuses in positive identification. J Forensic Sci 2005;50(1):18–22. 16. Stephan CN, Winburn AP, Christensen AF, Tyrrell AJ. Skeletal identification by radiographic comparison: blind tests of a morphoscopic method using antemortem chest radiographs. J Forensic Sci 2011;56 (2):320–32. 17. Riepert T, Ulmcke D, Schween F, Nafe B. Identification of unknown dead bodies by x-ray image comparison of the skull using the x-ray simulation program FoXSIS. Forensic Sci Int 2001;117:89–98. 18. de Ribeiro F. Standardized measurements of radiographic films of the frontal sinuses: an aid to identifying unknown persons. Ear Nose Throat J 2000;26:32–3. 19. Watamaniuk L, Rogers T. Positive personal identification of human remains based on thoracic vertebral margin morphology. J Forensic Sci 2010;55(5):1162–70. 20. Bunch AW, Fielding CG. The use of World War II chest radiograph in the identification of a missing-in-action U.S. Marine. Mil Med 2005;170 (3):239–42.

8

JOURNAL OF FORENSIC SCIENCES

21. Adams BJ, Maves RC. Radiographic identification using the clavicle of an individual missing from the Vietnam conflict. J Forensic Sci 2002;48 (2):369–73. 22. Scott AL, Congram D, Sweet D, Fonseca S, Skinner M. Anthropological and radiographic comparison of antemortem surgical records for identification of skeletal remains. J Forensic Sci 2009;55(1):241–4. 23. Spitz DJ. Identification of human remains. In: Spitz WU, Spitz DJ, editors. Medicolegal investigation of death. Springfield, IL: Charles C. Thomas, 2006;184–239. 24. http://quickfacts.census.gov/qfd/states/48/48199.html (accessed May 6, 2013). 25. http://www.hsc.unt.edu/departments/pathology_anatomy/dna/Forensics/ Initiative/Initiative.cfm (accessed May 6, 2013). 26. Zhao KD, Yang C, Zhao C, Stans AA, An KN. Assessment of noninvasive intervertebral motion measurements in the lumbar spine. J Biomech 2005;38:1943–6. 27. Reitman CA, Hipp JA, Nguyen L, Esses SI. Changes in segmental intervertebral motion adjacent to cervical arthrodesis. a prospective study. Spine 2004;29:E211–26. 28. Reitman CA, Mauro KM, Nguyen L, Ziegler J, Hipp JA. Intervertebral motion between flexion and extension in asymptomatic individuals. Spine 2004;24:2832–43. 29. Thompson MT, Conditt MA, Ismaily SK, Agarwal A, Noble PC. Brief report: validation of a system for automated measurement of knee laxity. Clin Biomech 2004;19:308–12. 30. Hummel RA. Image enhancement by histogram transformation. Comput Graph Image Process 1977;6:184–95. 31. North DO. An analysis of the factors which determine signal/noise discrimination in pulsed carrier systems. Princeton, NJ: RCA Labs, 1943 Princeton, NJ, Report No.: PTR-6C. 32. Kalman RE. A new approach to linear filtering and prediction problems. J Basic Eng 1960;82(1):35–45. 33. Dice LR. Measures of the amount of ecologic association between species. Ecology 1945;26(3):297–302. 34. Zou KH, Warfield SK, Bharatha A, Tempany CM, Kaus MR, Haker SJ, et al. Statistical validation of image segmentation quality based on a spatial overlap index. Acad Radiol 2004;11(2):178–89.

35. Jaccard P. The distribution of flora in the alpine zone. New Phytol 1912;11:37–50. 36. Wang Z, Bovik AC, Sheikh HR, Simoncelli EP. Image quality assessment: from error visibility to structural similarity. IEEE Trans Image Process 2004;13(4):600–12. 37. Peng HC, Long F, Ding C. Feature selection based on mutual information: criteria of max-dependency, max-relevance, and min-redundancy. IEEE Trans Pattern Anal Mach Intell 2005;27(8):1226–38. 38. Campbell JD. Edge structure and the representation of pictures [dissertation]. Columbia, MO: University of Missouri, 1969. 39. http://www.cdc.gov/nchs/nhanes/nhanesii.htm (accessed May 20, 2013). 40. Ratha NK, Connell JH, Bolle RM. Enhancing security and privacy in biometrics-based authentication systems. IBM Syst J 2001;40(3):614–34. 41. Kondo K, Morishita J, Katsuragawa S, Doi K. Development of an automated patient-recognition method for digital chest radiographs using edge-enhanced images. Jpn J Radiol Tech 2003;59(10):1277–84. 42. Morishita J, Katsuragawa S, Sasaki Y, Doi K. Potential usefulness of biological fingerprints in chest radiographs for automated patient recognition and identification. Acad Radiol 2004;11:309–15. 43. Christensen AM, Crowder CM. Evidentiary standards for forensic anthropology. J Forensic Sci 2009;54(6):1211–6. 44. Rogers TL, Allard TT. Expert testimony and positive identification of human remains through cranial suture patterns. J Forensic Sci 2004;49 (2):203–7. 45. Shamir L, Ling S, Rahimi S, Ferrucci L, Goldberg IG. Biometric identification using knee x-rays. Int J Biom 2009;1(3):365–70. Additional information and reprint requests: Sharon M. Derrick, Ph.D. Agency Coordinator/Forensic Anthropologist Harris County Institute of Forensic Sciences 1885 Old Spanish Trail Houston, TX 77054 E-mail: [email protected]