Forensic scientists put classic nose prediction rules to the test with CT scans

When a skeleton is found and DNA, fingerprints and dental records cannot deliver a name, forensic scientists sometimes turn to the skull itself, rebuilding a face from bone in the hope that a grieving relative or a missing-persons database will recognize it. The nose is one of the most prominent and recognizable features in this process, known as craniofacial identification, and yet the rules used to reconstruct it have rarely been put to rigorous independent test. A new study published in the International Journal of Legal Medicine by Yong Wei Xue and Carl N. Stephan of the Laboratory for Human Craniofacial and Skeletal Identification at the University of Queensland has now done exactly that, using high-resolution computed tomography scans of 92 Australian adults to measure how accurately three long-standing nose estimation models actually predict the faces of real people.

The three methods under scrutiny represent some of the most widely used guidelines in the field. The first is the so-called three-fifths rule, which holds that the width of the soft nose equals the maximum width of the bony nasal aperture divided by 0.6. The second, proposed by Davy-Jow and colleagues in 2012, claims that the contour of the superior nasal aperture mirrors the shape of the nasal tip in a one-to-one fashion when the head is rotated dorsally by about sixty degrees from the Frankfurt horizontal plane. The third is a revised version of the famous two-tangent method attributed to Mikhail Gerasimov, the Soviet pioneer of facial reconstruction, in which the tip of the nose, the pronasale, is located at the intersection of a line projected from the distal tip of the nasal bones and a second line projected from the floor of the nasal aperture.

The study’s raw material was a set of full-head CT scans of embalmed body donors to the University of Queensland anatomy program, comprising 53 males and 39 females aged between 51 and 100 years. The scans were acquired with a 16-slice Siemens Biograph helical CT at a slice thickness of just 0.75 millimeters, using settings of 130 kilovolts and 220 milliamperes-seconds. Decedents were deliberately chosen rather than living volunteers because the thin-slice protocol needed for such anatomical clarity would deliver radiation doses far above what is permissible for clinical scans of living people, and because donors remain perfectly still during acquisition, eliminating the blurring caused by breathing, heartbeat and muscle tone. Surface meshes of both skull and face were exported from the scans into Blender for measurement and analysis, with bone segmented at Hounsfield unit thresholds of 300 to 400 and soft tissue between minus 200 and minus 350.

Because the nose is an extremely pliable structure that loses its elastic recoil after death, the researchers carefully screened every subject for even the slightest compression or deformation of the nasal tissues. This screening meant that different subsets of the original 92 subjects were used to test each method: 63 individuals for the nose width rule, 31 for the nasal tip shape method, and 43 for the pronasale position technique. The authors argue that this conservative approach ensured that only the highest fidelity data entered each analysis, even though it reduced sample sizes.

The results for nose width were, on balance, encouraging but nuanced. The mean ratio between nasal aperture width and soft nose width across the 63 subjects came out at 0.614, remarkably close to the 0.6 value of the three-fifths rule, confirming that the guideline captures the central tendency of the population well. Mean errors were small, at 0.7 millimeters for males and 1.5 millimeters for females. However, the standard errors of the estimate were around 4 millimeters for both sexes, and the correlation between bony and soft tissue widths was a modest 0.32. In practical terms, this means that while the rule is accurate on average, individual predictions can be substantially wrong, with some people’s noses deviating widely from the population mean. The authors note that this finding echoes earlier work: Rynn and colleagues reported a mean error of only 0.4 millimeters in a 2010 CT study, while Ryu and colleagues found a correlation of 0.35 in South Koreans, closely matching the present result.

The nasal tip shape test delivered a more sobering verdict. Using geometric morphometric techniques, the researchers extracted 31 landmarks along both the bony curve of the superior nasal aperture and the soft tissue curve of the nasal tip for each of the 31 screened subjects, then compared the curves after Procrustes registration, which removes differences in position, rotation and scale. The resulting Procrustes distance of 0.24 indicates a moderate but real shape mismatch. Mean point-to-point differences between the curves averaged 4.0 millimeters, and visual inspection of the mean curves showed that the hard tissue contour is considerably more peaked at its apex than the soft tissue curve, which is flatter and more gently rounded. This is the first empirical cross-validation of the Davy-Jow method, and the authors conclude that the assumption of a direct one-to-one correspondence between the bony aperture and the nasal tip outline should be treated with considerable skepticism.

The pronasale position test focused on Gerasimov’s revised two-tangent method, incorporating clarifications published by Ullrich and Stephan in 2011 based on archival research into Gerasimov’s actual practice. These clarifications revealed that Gerasimov used only the very last 1 to 2 millimeters of the nasal bone to set the upper tangent, not the last third as his published instructions stated, and that the lower tangent was drawn not from the nasal spine but from the floor of the nasal aperture beside the vomer bone. In a methodological advance, the Queensland team abandoned the traditional practice of eyeballing these tangent lines and instead used Blender’s mesh editing tools to select triangular faces of the skull tessellation, calculate the mean surface normal for each region of interest, and attach tangents to the resulting quantitatively defined planes. This removed subjective judgment from a step that has always been performed by visual estimation.

Even with this precision, the two-tangent method showed substantial errors. The mean x-coordinate error was 2.2 millimeters, indicating that the estimated nose tip sits too far forward, while the mean y-coordinate error was minus 6.2 millimeters, indicating it sits too far down. The Euclidean distance between estimated and true pronasale positions averaged 8.3 millimeters, with a standard error of the estimate of 10 millimeters. These figures closely replicate those of Maltais Lapointe and colleagues, who in 2016 validated the revised interpretation using CT scans of 66 subjects and concluded that the error of any version of Gerasimov’s method is too high for reliable identification work. The present study also revealed large intra-observer variability in placing the estimated pronasale, with a technical error of measurement of 4.0 millimeters on the x-axis, underscoring the method’s limited repeatability even when tangents are defined numerically.

Rather than simply discarding the flawed methods, the authors propose a statistical remedy known as calibration-in-the-large. By applying the sign-reversed mean errors as correction factors, systematic biases in the estimation models can be removed. For the two-tangent method, translating the estimated pronasale by minus 2.2 millimeters in x and plus 6.2 millimeters in y reduced the standard error of the Euclidean distance from 10.0 to 2.5 millimeters in the study’s own training data, a fourfold improvement, and from 10.1 to 8.3 millimeters in the independent dataset of Maltais Lapointe and colleagues. Similar corrections of minus 0.7 millimeters for males and minus 1.5 millimeters for females were proposed for the nose width rule. The authors caution that these calibrations are derived from training data and may suffer some overfitting, so their true utility must be confirmed in future out-of-group samples. Nonetheless, the study marks an important step toward a more honest accounting of facial approximation: a field that can now quantify, in millimeters, exactly how much its oldest rules bend the truth, and begin to correct them.

Subject of Research: CT-based validation of nose width, nasal tip shape and pronasale position estimation models for craniofacial identification

Article Title: Nose-skull relationships for craniofacial identification: computed tomography (CT) assessment of nose width, nasal bridge shape & pronasale position

Article References: Xue, Y. W., & Stephan, C. N. (2026). Nose-skull relationships for craniofacial identification: computed tomography (CT) assessment of nose width, nasal bridge shape & pronasale position. International Journal of Legal Medicine. https://doi.org/10.1007/s00414-026-04009-3

Image Credits: AI Generated

DOI: 10.1007/s00414-026-04009-3

Keywords: craniofacial identification, facial approximation, forensic anthropology, nose width, pronasale position, nasal aperture, computed tomography, Gerasimov two-tangent method, geometric morphometrics, facial reconstruction, calibration-in-the-large, International Journal of Legal Medicine