The research, led by Mauranne Lievens of the Department of Oral & Maxillofacial Surgery at Ghent University Hospital, together with Wim Van Paepegem, Tom Goffin, Geert Villeirs and Renaat Coopman, addresses a long-standing gap in the validation of hospital 3D printing programs. Earlier validation studies typically relied on geometric calibration artifacts, standardized test shapes, or simplified phantom objects—objects that behave well on a build plate but share little with the irregular, thin-walled, undercut-rich anatomy of a human skull. The Ghent team instead designed the study around highly accurate anatomical models, deliberately mimicking the actual workflow a hospital follows when a surgeon requests a patient-specific replica for preoperative planning, implant shaping, or patient counseling.

Accuracy in this context is not a single number but a composite of two distinct properties. The first is trueness, which measures how closely a printed model matches the digital reference design—the deviation between the physical object and the source file that clinicians use. The second is precision, which captures reproducibility: how consistent a printer is when producing the same model repeatedly within a single build (intra-build variability), and how consistent it is across separate print jobs run at different times (inter-build variability). Both dimensions matter clinically. A printer that faithfully reproduces anatomy once but drifts between builds is as problematic as one that is consistently wrong. The team quantified both using root mean square (RMS) error, a standard metric that aggregates surface deviation across the entire model rather than relying on a handful of caliper measurements at convenient points.

The statistical design of the study is one of its most notable contributions. Because the dataset comprised repeated prints of multiple anatomical models at multiple timepoints, the observations were inherently clustered—prints from the same build and models from the same printer are not independent of one another. Conventional statistical tests that ignore this structure can overstate significance. To handle the complexity, the researchers employed a Linear Mixed Model (LMM), a framework widely used in agriculture, ecology and clinical trials but, according to the authors, never before applied in 3D printing accuracy research. The mixed model allowed the team to evaluate fixed effects—printer type, anatomical model, and comparison type—while correctly accounting for the internal clustering of repeated measurements, yielding a statistically robust picture of where true differences lie.

The headline result is that every technology tested achieved accuracy within clinically acceptable limits, an encouraging finding for hospitals that have invested in lower-cost equipment. But the ranking was clear and statistically significant. Material Jetting, which builds objects by jetting tiny droplets of photopolymer in extremely thin layers, achieved a mean RMS error of 67 micrometers—the finest accuracy in the study. Stereolithography, which cures liquid resin layer by layer with a light source, followed at 109 micrometers. Fused Filament Fabrication, the thermoplastic extrusion process behind most affordable desktop printers, came in at 130 micrometers. In practical terms, these differences span roughly the width of one to two human hairs, yet in surgery that margin can matter: when a model is used to pre-bend titanium plates or rehearse the osteotomy of a complex orbital or mandibular reconstruction, sub-millimeter deviations propagate directly to the operating table.

Just as important as the ranking was what the Linear Mixed Model revealed about the drivers of error. Printer type was a significant fixed effect, confirming that the technology choice itself—not random variation—explains much of the accuracy difference. But anatomical complexity also exerted a significant influence, meaning that the shape being printed matters nearly as much as the machine printing it. Models with delicate facial structures, thin bony walls and deep undercuts stressed each technology differently: resin- and jetting-based photopolymer processes handled fine detail more gracefully, while extrusion-based FFF showed characteristic stair-stepping and filament-path limitations on curved, intricate surfaces. This interaction underscores why calibration artifacts alone cannot predict clinical performance; a flat test coupon tells a hospital little about how a printer will cope with a zygomatic arch or a thin orbital floor.

The study’s framing around the Medical Device Regulation gives it particular weight for European hospitals. MDR 2017/745 classifies many in-house manufactured products as devices and requires health institutions to demonstrate that their production, under the responsibility of a single legal manufacturer, meets safety and performance expectations—including verification of accuracy appropriate to the intended clinical use. Point-of-care 3D printing has often grown organically inside surgical departments, with validation practices varying wildly from one hospital to the next. By defining accuracy through trueness and precision, quantifying it with RMS error against digital references, and analyzing it with a defensible statistical model, the Ghent team has effectively published a template for the kind of documented, reproducible validation that regulators and hospital quality officers can build upon.

For hospital administrators weighing procurement decisions, the results offer a nuanced cost-benefit picture. Material Jetting’s superior fidelity comes with higher machine and material costs, and photopolymer processes require resin handling and post-processing protocols. FFF remains the most affordable and accessible option, and the study confirms its output is still clinically acceptable—suggesting it can retain a role for models where extreme fidelity is not essential, such as patient education or approximate planning. SLA sits between the two, offering strong performance at moderate cost. The finding that all three cleared the clinical acceptability bar means the choice can legitimately be driven by the intended application, budget and workflow, rather than by fear that cheaper technology is automatically unsafe. For applications demanding the highest anatomical fidelity within the MDR framework, however, the evidence now points decisively toward Material Jetting.

The research also carries implications well beyond craniomaxillofacial surgery. In-house and point-of-care 3D printing is expanding rapidly into orthopedics, cardiothoracic surgery, dental applications and surgical training, and virtually every one of those programs faces the same questions: how accurate is our printer, how repeatable is it, and can we prove it? The study’s methodological chain—realistic anatomical models, digitized reference comparisons, RMS-based quantification of trueness and precision, and mixed-model statistics that respect the structure of repeated manufacturing data—offers a portable blueprint for any institution seeking to validate its own fleet. The authors note it is the first application of a Linear Mixed Model in this field, and if the approach is adopted broadly, it could standardize how hospitals across jurisdictions demonstrate regulatory compliance. Published open access in the journal 3D Printing in Medicine, with no specific external funding and no declared competing interests, the study was reviewed with the involvement of the Ghent University Hospital Medical Ethics Committee, receiving official approval on July 2, 2024, and received on 16 September 2025, accepted on 16 August 2026, and published on 11 September 2026.

Subject of Research: Comparative accuracy of FFF, SLA and Material Jetting 3D printing for in-house anatomical models in craniomaxillofacial surgery

Article Title: Clinical relevance of accuracy in in-house 3D printing in craniomaxillofacial surgery: a comparative study of FFF, SLA, and MJ technologies

Article References: Lievens, M., Van Paepegem, W., Goffin, T., Villeirs, G., & Coopman, R. (2026). Clinical relevance of accuracy in in-house 3D printing in craniomaxillofacial surgery: a comparative study of FFF, SLA, and MJ technologies. 3D Printing in Medicine. https://doi.org/10.1186/s41205-026-00344-8

Image Credits: AI Generated

DOI: 10.1186/s41205-026-00344-8

Keywords: 3D printing, craniomaxillofacial surgery, anatomical models, FFF, SLA, Material Jetting, Medical Device Regulation, accuracy validation, Linear Mixed Model, point-of-care manufacturing, Clinical, relevance

The scale of the underlying dataset is what makes the achievement notable. Between the 2021 and 2024 cropping seasons, the researchers assembled 22 separate multi-environment trials involving 1,035 bread wheat genotypes, including ten released check varieties such as Abay, Kakaba, Pavon-76 and Kingbird. The trials spanned seven locations—Arsi Negelle, Kulumsa, Dhera, Gorro, Enawari, Alem Tena and Zeway—that together represent the diverse low-altitude agro-ecologies where Ethiopia’s wheat future increasingly depends. Because not every genotype was grown at every site, the data were deliberately unbalanced, a situation that defeats classical analysis but is handled naturally by the mixed model framework the team employed.

At the heart of the approach lies a statistical distinction that most field trials gloss over: the difference between what a plant’s genes make it capable of and what the environment lets it express. The researchers fitted factor analytic mixed models using restricted maximum likelihood, treating genotypes as random effects and estimating their performance through Best Linear Unbiased Predictions, or BLUPs. Unlike simple averages, BLUPs ‘shrink’ extreme values toward the overall mean, damping the influence of lucky plots and noisy sites. Spatial models simultaneously captured local, extraneous and global field trends within each trial, so that soil gradients and micro-climate patches did not masquerade as genetic superiority.

The results revealed just how variable the testing network was. Genetic variance for grain yield ranged from a negligible 0.02 to a substantial 1.12 across trials, while error variance spanned 0.13 to 0.75. Sites such as 22BWPEKU, 22BWNEKU, 24BWOPNEKU and 21BWOEAN showed high genetic variance, marking them as powerful ‘discriminating’ environments where the true differences among genotypes could shine through. Other locations, including 24BWPNEZW, 22BWNEGR and 22BWNEAT, exhibited such low genetic signal that the team effectively excluded their BLUPs from final selection decisions—a quality-control step that prevented noisy environments from diluting the breeding index.

Heritability estimates told a parallel story. Days to heading proved remarkably stable, with values between 70.25 and 98.63 percent, and hectoliter weight and thousand kernel weight also showed robust heritability, ranging from roughly 56 to 96 percent. Grain yield, by contrast, swung between 22.92 and 92.71 percent, and plant height between 4.74 and 91.89 percent, underscoring that yield is the trait most hostage to environmental whims. This pattern has practical consequences: breeders can select confidently for maturity and grain characteristics in fewer locations, but yield stability demands a genuine multi-environment strategy.

To map that strategy, the team used dendrograms and heat maps built from the genetic correlation matrices of the fitted factor analytic models. The 22 environments resolved into nine genotype-by-environment clusters for grain yield, with clusters C1 through C7 forming the core selection framework and clusters C8 and C9, which showed weak genetic correlations with the rest, analyzed independently. Cluster 3, dominated by Arsi Negelle trials, recorded the highest mean grain yields, while Cluster 6, centered on Kulumsa, emerged as a highly sensitive site for separating elite lines from average ones. Days to heading, the most stable trait, grouped into only two clusters, while grain yield’s nine clusters confirmed it as the most environmentally sensitive character measured.

The factor analytic models themselves proved remarkably efficient at capturing the underlying genetic architecture. In twelve environments the three-factor model explained nearly 100 percent of the total genetic variance, with Factor 1 accounting for up to 99.84 percent in trial 22BWOEKU and Factor 2 reaching 99.48 percent in 24BWPNEZW. A handful of outlier trials—six in total—resisted the model, likely reflecting extreme weather events during critical growth stages that decoupled those sites from the broader network. The researchers caution that such exceptionally high explained variance should be interpreted carefully, but the overall pattern confirms that factor analytic structures offer a parsimonious, computationally robust approximation to fully unstructured genotype-by-environment covariance.

When the selection index of averaged BLUPs was applied across the correlated clusters, five genotypes rose to the top: EBW192940, EBW212724, EBW180175, EBW212777 and EBW212106, alongside the check variety Hawi, all exceeding 6.55 tonnes per hectare in predicted mean yield and outperforming the standard check Asgori. More than 60 percent of the 1,035 evaluated genotypes averaged above 4.5 tonnes per hectare, signaling a deep pool of promising germplasm. The standout, EBW192940, was officially released in 2025 as the variety ‘Dhera’ for low-altitude wheat-growing areas, a concrete deliverable that validates the entire analytical pipeline from nursery screening through national variety trials.

The implications extend well beyond one variety. Ethiopia cultivates roughly 2.1 million hectares of wheat, yet national productivity remains below the attainable potential of about 5 tonnes per hectare, a gap driven largely by the scarcity of high-yielding, stable varieties adapted to diverse ecologies and by persistent biotic and abiotic stresses. By identifying which testing sites genuinely discriminate among genotypes, which clusters of environments share a common ranking, and which lines hold their performance across years and locations, the linear mixed model framework gives breeders a rational map for deploying resources. The authors note limitations—the study covered a constrained set of locations and seasons, focused mainly on grain yield, and did not incorporate quality, disease resistance or farmer preference traits—but the direction is clear. As breeding programs across the developing world grapple with increasingly erratic climates, the Ethiopian wheat experience suggests that the fastest route to climate-resilient harvests may run not through new genes alone, but through smarter statistics applied to the trials breeders are already running.

The choice of experimental design within each trial deserves attention because it underpins the reliability of everything that followed. The researchers employed row-column Alpha-lattice designs alongside partially replicated designs, known in breeding circles as p-rep arrangements. In a p-rep design, rather than repeating every genotype two or three times across a field, breeders replicate only a subset of promising lines while others appear just once. This allows the same amount of land to carry substantially more genetic diversity per trial, which is precisely what a program screening over a thousand genotypes requires. The trade-off is that single-plot entries carry more uncertainty, but the mixed model framework converts that scattered replication into precise predictions by borrowing strength across the entire network of trials.

The partially replicated approach also reflects a broader shift in how breeding programs manage scarce resources. Traditional designs that fully replicate every entry consume plot space, seed, labor and budget in proportion to genotype numbers, which grows every cycle as new crosses are advanced. By concentrating replication where it matters most—among the lines most likely to progress toward release—a program can evaluate far more germplasm per season. The Ethiopian low-altitude program’s adoption of this design, combined with spatial adjustment, illustrates how statistical sophistication and practical field logistics reinforce one another rather than competing for the same resources.

Another dimension worth highlighting is the treatment of heterogeneous error across environments. In a network spanning seven locations over four seasons, the assumption that measurement noise is identical everywhere is untenable. Rainfall patterns, soil fertility gradients, disease pressure and management practices differ from site to site, and even within a single field the error structure can vary. The linear mixed model framework allowed the analysts to estimate separate error variances for each environment, as the reported range from 0.13 to 0.75 makes evident. Had a single pooled error term been used, trials with inherently noisy conditions would have dragged down the precision of every other environment’s estimates, and genuinely superior genotypes tested at discriminating sites might have been overlooked.

The classification of environments into groups with similar genotype rankings also carries forward-looking value for breeding strategy. When two locations show strong positive genetic correlation, a genotype performing well at one will predictably perform well at the other, meaning duplicated testing at both sites adds little information relative to its cost. Conversely, environments that correlate weakly or negatively with the broader network represent distinct selection targets that may require dedicated breeding efforts. By quantifying these relationships, the factor analytic models effectively provide a data-driven rationale for deciding where future trials should be conducted, how many testing locations are genuinely needed, and whether the target population of environments should be subdivided into separate product profiles for variety development.

Finally, the trajectory from statistical prediction to an officially released variety demonstrates the end-to-end function of a national breeding pipeline. The five elite lines identified through averaged BLUPs did not remain abstractions on a spreadsheet; the leading genotype entered the formal verification and release process and emerged as ‘Dhera’, named for one of the very locations in the testing network. That close connection between analytical output and tangible farmer-facing outcomes is what distinguishes a functioning variety development system, and it offers a template that other crop programs in comparable agro-ecologies can adapt to their own multi-environment trial data.

Subject of Research: Linear mixed model analysis of multi-environment trial data for bread wheat genotype selection in low-altitude Ethiopia

Article Title: Analysis of multi-environment trial (MET) Data using linear mixed model for bread wheat genotypes across low altitude areas of Ethiopia

Article References: Asefa, B., Zegeye, H., Geleta, N., Sime, B., Solomon, T., Dabi, A., Alemu, G., Dhuga, R., Asnake, D., Delesa, A., Zewdu, D., & Getamesay, A. (2026). Analysis of multi-environment trial (MET) Data using linear mixed model for bread wheat genotypes across low altitude areas of Ethiopia. BMC Agriculture, 2(1), Article 27. https://doi.org/10.1186/s44399-026-00053-x

Image Credits: AI Generated

DOI: 10.1186/s44399-026-00053-x

Keywords: bread wheat, multi-environment trials, linear mixed model, factor analytic model, BLUP, genotype-by-environment interaction, Ethiopia, plant breeding, grain yield, heritability, variety release, agricultural statistics