Detecting temporal asymmetry after epilepsy surgery: a 3D MRI-based comparative outcome study of clinicians and lay observers

Introduction

Perfect symmetry does not exist in humans; instead, craniofacial asymmetry is viewed as a continuum, from normal craniofacial fluctuating asymmetry in healthy individuals (Crins-de Koning et al., 2025) to pronounced congenital, developmental, and acquired deformities (Cheong & Lo, 2011; Wenger, Gallagher & Bhoj, 2019). Among craniofacial subunits, the frontally visible temporal region plays a key role in facial symmetry, as its contour—positioned above the zygomatic arch and behind the lateral orbital rim—defines the upper facial width (Vaca et al., 2017; Shay et al., 2022). Unilateral changes in this area can result in noticeable asymmetry (Santiago et al., 2018; Thiensri, Limpoka & Burusapat, 2020).

In this setting, resective epilepsy surgery via the pterional approach is a notable contributor to asymmetrical temporal deformities, due to epilepsy’s global prevalence (Yang et al., 2024), the need for surgery in drug-resistant epilepsy (Pang et al., 2025), and potential postoperative complications (Yasuda et al., 2010; Gonzalez-Martinez, 2025). These deformities often lead to appearance-related preoccupation, disrupting quality of life, self-perception, daily functioning, and social interactions (Klassen et al., 2021), underscoring the need for greater clinical and research attention.

The amount of craniofacial asymmetry reported in both healthy individuals and affected patients varies depending on the assessment method used (Nguyen et al., 2024; Harripershad et al., 2025; Lin et al., 2025; Nishimura et al., 2025). Panel assessment, a common indirect method for evaluating craniofacial asymmetry, relies on subjective judgments from lay observers and clinicians (Zhu, Jayaraman & Khambay, 2016; Tan et al., 2021), with clinicians generally considered more sensitive, though studies report mixed findings on their relative accuracy (Lee, Dumrongwongsiri & Lo, 2019; Tan et al., 2021; Zhang et al., 2023). This approach is widely used to detect visible asymmetries and establish recognition thresholds across various craniofacial subunits, which differ by region (Wang et al., 2017; Lee, Dumrongwongsiri & Lo, 2019). For instance, occlusal canting under 3° often goes unnoticed, while deviations over 4° are detected by 90% of observers (Padwa, Kaiser & Kaban, 1997). However, formal data specific to the temporal region is lacking.

Imaging modality can also influence asymmetry-focused analyses (Lo & Lin, 2023). Subjective and objective craniofacial assessments are limited when based on two-dimensional images, which are prone to distortion, magnification errors, and lack of depth representation (Hsu et al., 2020). The rise of three-dimensional (3D) imaging has improved accuracy, offering undistorted, lifelike representations of the complex craniofacial structure (Wu et al., 2019). Among various 3D reconstruction methods, magnetic resonance imaging (MRI) offers non-invasive, radiation-free, high-contrast, high-resolution soft tissue imaging without visual distractions such as skin color, hair, or facial features that could bias perception (Vander Pluym et al., 2007; Villavisanis et al., 2024). MRI is essential for diagnosing and managing epilepsy, offering high-quality data storage (Biagioli et al., 2025). Neuroimaging repositories like Brazilian Institute of Neuroscience and Neurotechnology (BRAINN) house MRI datasets that have supported epilepsy research and can also contribute to craniofacial studies using images originally acquired for other neuroscience purposes (De Souza et al., 2020; Giacomini et al., 2020).

Recent systematic reviews confirm that no standardized method for analyzing the temporal region has been established (Wang et al., 2017; Gonçalves et al., 2021; Shay et al., 2022; Nasim et al., 2024). Although several grading systems have been proposed, their validity and consistency remain insufficiently examined (Vaca et al., 2017; Wang et al., 2017; Choi et al., 2018; Kim et al., 2018; Huang et al., 2018; Laloze et al., 2019; Gonçalves et al., 2021; Shay et al., 2022; Nasim et al., 2024). Moreover, reliable assessments of unmanipulated 3D images for detecting temporal asymmetry after epilepsy surgery remain lacking. Better understanding how laypeople and clinicians perceive such asymmetry could improve deformity detection, treatment planning, and research design.

Using MRI-derived 3D craniofacial soft-tissue image datasets from adult patients treated with resective surgery for temporal lobe epilepsy, this study aimed to assess the subjective recognition of temporal asymmetry by comparing evaluations from trained clinicians and lay observers. Additionally, the study aimed to determine the threshold value at which a temporal asymmetry is clinically recognized as abnormal.

Materials & Methods

This methodological cross-sectional investigation (Fig. 1) analyzed a standardized craniofacial image dataset obtained from the BRAINN database following approval by the Institutional Review Board (CAAE: 93412318.0.0000.5404; University of Campinas (UNICAMP), Campinas, São Paulo, Brazil) and in compliance with the Declaration of Helsinki. All patients provided written informed consent prior to undergoing MRI scans. The study included MRI-derived images from Brazilian adult patients (n = 96; 53 females; aged 42.6 ± 11.2 years; 58 left-side temporal lobe epilepsy) who consecutively underwent unilateral resective surgery via a modified pterional approach for drug-resistant epilepsy between 2016 and 2022, with a minimum postoperative period of 12 months. The 12-month postoperative evaluation was selected to ensure complete scar maturation, resolution of soft-tissue swelling, and stabilization of surgery-induced change (Bond et al., 2008). All included patients had a confirmed diagnosis of temporal lobe epilepsy, established by one of two experienced epileptologists at the Epilepsy Clinic, UNICAMP, Brazil. Exclusion criteria (n = 11) included the presence of other pathological conditions affecting the head or face (e.g., dento-skeletofacial deformities, prior reconstructive procedures, craniofacial trauma, or surgery) and incomplete datasets.

The imaging databank was compiled for diagnostic and follow-up purposes, with written patient consent (Alvim et al., 2016; De Souza et al., 2020; Giacomini et al., 2020). Under standardized conditions previously described by our team, all pre- and postoperative MRI images were acquired (3 Tesla Achieva-Intera Philips^®: T1-weighted images with isotropic voxels of one mm, slice thickness one mm, no gap, flip angle = 8°, TR = 7.0 ms, TE = 3.2 ms, matrix = 240 × 240, and field of view = 240 × 240) (Fonseca et al., 2012; Coan et al., 2014; Campos et al., 2015; Alvim et al., 2016; De Campos et al., 2016) and preprocessed using a multi-software workflow, with intensity inhomogeneity corrected by the N4 bias field correction algorithm in 3D Slicer and intensity normalization achieved through global scaling and alignment to tissue probability maps during spatial normalization in SPM8 (Ashburner & Friston, 2005; Ashburner, 2007; Avants, Nick & Gang, 2009; Tustison et al., 2010; Fedorov et al., 2012; Chakravarty et al., 2011; Madan, 2015; Alvim et al., 2016; Yushkevich et al., 2016; Villavisanis et al., 2024). Postoperative MRI images were rigidly co-registered to the preoperative images using the General Registration (BRAINS) module in 3D Slicer (Johnson, Harris & Williams, 2007; Kikinis, Pieper & Vosburgh, 2014), ensuring anatomical correspondence for asymmetry assessment without altering local tissue morphology (Friston et al., 1995; Ashburner & Friston, 1997; Ashburner & Friston, 2005).

A combination of advanced image analysis and engineering software was utilized for image processing and measurement: Avizo (FEI, Mérignac, France), Geomagic (3D Systems, Rock Hill, S.C.), and SimPlant O&O (Materialize, Leuven, Belgium). All anatomical landmarks, head orientation reference frames, measurement reference planes, severity stratification levels (I to IV), potential cutoff value for detecting asymmetrical deformities of the temporal region, subjective panel assessments (including 3D image frontal views, task time, and binominal scale system), and objective linear measurement (temporal soft tissue thickness difference in percentage) were standardized according to prior studies (Padwa, Kaiser & Kaban, 1997; Vander Pluym et al., 2007; Stephan & Devine, 2009; Kim et al., 2018; Wu et al., 2019; Hsu et al., 2020; Wan, Tsai & Lo, 2021; Morandi et al., 2022; Kurniawan et al., 2024).

Objective analysis

Using predefined anatomical landmarks and structures of interest—including the lateral orbital rim, temporal bone surface, and external skin surface—identified and interactively verified in the axial, coronal, and sagittal imaging planes, objective measurements of temporal thickness differences (Fig. 2A) were collected from each preoperative and postoperative MRI scan (Vander Pluym et al., 2007; Stephan & Devine, 2009; Bu et al., 2010; Welling et al., 2015; Wysong et al., 2013; Wysong et al., 2014; Kim et al., 2018; Morandi et al., 2022).

The degree of temporal asymmetry was calculated as the percentage difference in thickness between the right and left temporal sides (or between the operated and non-operated sides) using the formula: (longer side−shorter side)/longer side×100. A total of 192 MRI scans (96 pre-epilepsy surgery and 96 post-epilepsy surgery) were linearly ranked based on incremental asymmetry values. Temporal asymmetry was subsequently categorized into four severity levels (I, II, III, and IV), each containing 48 images, ranging from the minimum to the maximum asymmetry percentage values.

Using the temporal thickness difference data, a previously described but not validated cutoff value of 10% for temporal asymmetry (Kim et al., 2018) was applied to classify included preoperative and postoperative images into “no visible temporal asymmetry” (≤cut-off value) versus “at risk of clinically obvious temporal asymmetry deformity” (>cut-off value). This cutoff value was further utilized for comparative analysis of subjective asymmetry detection by various observer categories, including trained clinicians and untrained observers, as well as subcategories such as family members, laypeople, clinical observers, and surgical observers.

Results

Objective analysis

For objective measurements, intra- and inter-examiner reliability ranged from moderate to excellent. Regarding the error magnitude statistic, temporal thickness measurements demonstrated good precision for the REM parameter (Table S1).

Most MRI craniofacial images (61.5%; P < 0.001) exhibited temporal thickness differences of less than 10% (Table 1; Fig. 3). Across asymmetry severity levels, level I showed the lowest temporal thickness values, followed by levels II through IV (all P < 0.001). Asymmetry levels I and II had no temporal thickness differences exceeding 10%, while level III included both values above and below 10% (54.2% and 45.8% of images, respectively; P > 0.05). Level IV consisted exclusively of images with values above 10% (Table 1; Fig. 4).

Table 1:

Temporal thickness difference distribution across asymmetry levels.

Temporal asymmetry level	Temporal thickness difference			P-value
	Overall	≤10%	>10%
Level I–IV
n (%)	192 (100)	118 (61.46)	74 (38.54)	<0.001*
Temporal thickness difference (%) m ± sd (range)	11.46 ± 9.77 [1.12–46.56]	4.87 ± 2.08 [1.12–9.90]	21.96 ± 7.80 [10.19–46.56]	<0.001*
Level I
n (%)	48 (25)	48 (100)	0 (0)	–
Temporal thickness difference (%) m ± sd (range)	2.93 ± 0.99 [1.12–4.12]	2.93 ± 0.99 [1.12–4.12]	0 ± 0 [0–0]	–
Level II
n (%)	48 (25)	48 (100)	0 (0)	–
Temporal thickness difference (%) m ± sd (range)	5.35 ± 0.61 [4.13–6.52]	5.35 ± 0.61 [4.13–6.52]	0 ± 0 [0–0]	–
Level III
n (%)	48 (25)	22 (45.83)	26 (54.17)	0.541
Temporal thickness difference (%) m ± sd (range)	11.52 ± 3.54 [6.67–17.19]	8.08 ± 1.08 [6.67–9.90]	14.43 ± 1.83 [10.19–17.19]	<0.001*
Level IV
n (%)	48 (25)	0 (0)	48 (100)	–
Temporal thickness difference (%) m ± sd (range)	26.04 ± 6.66 [18.39–46.56]	0 ± 0 [0–0]	26.04 ± 6.66 [18.39–46.56]	–
P-value	<0.001*,**	–	–	–

DOI: 10.7717/peerj.20201/table-1

Notes:

–, no data for comparison purpose.

n, number of craniofacial images.

%, percentage.

m, mean.

sd, standard deviation.

Preop, preoperative.

Postop, postoperative.

*Significant after applying the Bonferroni correction.

**Level I < Level II < Level III < Level IV, with p < 0.001 for all pairwise comparisons.

Postoperative images exhibited a significantly (P < 0.001) greater temporal thickness difference than preoperative images (Table 2). No preoperative image had a temporal thickness difference greater than 10% (ranging from 1.12% to 9.23%), while postoperative images displayed a broader range of temporal thickness differences, from 1.99% to 46.56% (Table 2; Fig. 4). Both preoperative and postoperative images were distributed across asymmetry severity levels I to III, while only postoperative images were categorized as level IV (all P < 0.001) (Table 2). Subanalyses revealed no significant influence (all P > 0.05) of the tested parameters (age, sex, and surgical side) on temporal thickness differences, considering preoperative/postoperative image types, asymmetry cutoff, and severity levels.

Table 2:

Temporal thickness difference distribution across asymmetry levels and preoperative/postoperative image types.

Temporal asymmetry level	Temporal thickness difference
	Overall			≤10%			>10%
	Preop	Postop	P-value	Preop	Postop	P-value	Preop	Postop
Level I–IV
n (%)	96 (100)	96 (100)	–	96 (100)	22 (22.92)	<0.001*	–	74 (77.08)
Temporal thickness difference (%) m ± sd (range)	4.60 ± 1.92 [1.12–9.23]	18.32 ± 9.64 [1.99–46.56]	<0.001*	4.60 ± 1.92 [1.12–9.23]	6.09 ± 2.33 [1.99–9.90]	0.013	–	21.96 ± 7.80 [10.19–46.56]
Level I
n (%)	42 (87.5)	6 (12.5)	<0.001*	42 (87.5)	6 (12.5)	<0.001*	–	–
Temporal thickness difference (%) m ± sd (range)	2.85 ± 1.00 [1.12–4.12]	3.51 ± 0.70 [1.99–4.03]	0.139	2.85 ± 1.00 [1.12–4.12]	3.51 ± 0.77 [1.99–4.03]	0.139	–	–
Level II
n (%)	40 (83.3)	8 (16.7)	<0.001*	40 (83.3)	8 (16.7)	<0.001*	–	–
Temporal thickness difference (%) m ± sd (range)	5.35 ± 0.61 [4.13–6.52]	5.35 ± 0.59 [4.46–6.24]	0.989	5.35 ± 0.61 [4.13–6.52]	5.35 ± 0.63 [4.46–6.24]	0.989	–	–
Level III
n (%)	14 (29.2)	34 (70.8)	<0.001*	14 (29.2)	8 (16.7)	0.131	–	26 (54.1)
Temporal thickness difference (%) m ± sd (range)	7.69 ± 0.94 [6.67–9.23]	13.09 ± 2.95 [7.22–17.19]	<0.001*	7.69 ± 0.94 [6.67–9.23]	8.76 ± 1.02 [7.22–9.90]	0.020	–	14.43 ± 1.83 [10.19–17.19]
Level IV
n (%)	–	48 (100)	–	–	–	–	–	48 (100)
Temporal thickness difference (%) m ± sd (range)	–	26.04 ± 6.59 [18.39–46.56]	–	–	–	–	–	26.04 ± 6.66 [18.39–46.56]
P-value	–	<0.001*,**	–	–	–	–	–	–

DOI: 10.7717/peerj.20201/table-2

Notes:

–, no data for comparison purpose.

n, number of craniofacial images.

%, percentage.

m, mean.

sd, standard deviation.

Preop, preoperative.

Postop, postoperative.

*Significant after applying the Bonferroni correction.

**Level I < Level II < Level III < Level IV, with p < 0.001 for all pairwise comparisons.

Discussion

Currently, there is no gold standard for identifying temporal region asymmetry (Wang et al., 2017; Laloze et al., 2019; Shay et al., 2022; Nasim et al., 2024). To address this, instead of computer-generated asymmetrical models (e.g., chimeras or blends) (Naini et al., 2012a), we used unaltered 3D craniofacial models from surgically treated patients, preserving the natural clinical context and reflecting real-world epilepsy care in a Brazilian population of mixed ethnic backgrounds (Giacomini et al., 2020; Secolin et al., 2021). By applying appropriate statistical tests and established cut-off values (Padwa, Kaiser & Kaban, 1997; Rhodes et al., 2005; Chatrath et al., 2007; Kottner et al., 2011; Carvalho et al., 2012; Jackson et al., 2013; Lewis, 2017; Chou et al., 2019; Wan, Tsai & Lo, 2021), both the subjective panel assessments and the objective measurements of temporal soft tissue thickness differences met criteria for reliability and precision, supporting the robustness of the data collected for the temporal region.

Our analyses showed that 3D craniofacial images with prior temporal surgery exhibited greater asymmetry than those without surgical intervention, and perfect symmetry was absent in all images. These results support the presence of craniofacial asymmetry across the population (Chou et al., 2019; Crins-de Koning et al., 2025) and indicate that surgery-induced asymmetry is more clinically noticeable (Cheong & Lo, 2011). Our data reflect natural variability in the temporal region under real-world conditions, providing a useful reference for future studies on both natural and surgical asymmetry. To broaden these findings, future MRI-based research could include diverse ethnic groups and patients undergoing other temporal surgeries (e.g., for trauma, aneurysms, tumors) and assess additional asymmetry indices. Studies might also examine the influence of underlying brain and cranial structures on craniofacial asymmetry (Marečková et al., 2013; Kong et al., 2018).

In our study, subjective judgments of temporal asymmetry followed a consistent pattern: larger temporal thickness differences associated with higher detection rates. Both trained and untrained observers performed similarly for severe asymmetry (level IV), but detection rates varied significantly for milder asymmetries (levels I–III), with laypeople showing the lowest rates, followed by family members, while clinical and surgical observers demonstrated higher accuracy. Overall, lay observers detected asymmetry less frequently than clinicians.

Consistent with previous studies (Kokich, Kokich & Kiyak, 2006; An et al., 2014), specialized clinicians demonstrated superior recognition of asymmetry, likely due to their ability to apply clinical skills to subtle asymmetries. The clinicians in our study—craniofacial plastic surgeons and dermatologists experienced in treating appearance-altering craniofacial conditions—possess extensive expertise in managing temporal region asymmetries, including the use of fillers and grafts for reconstructive and aesthetic purposes. Unlike untrained observers, clinicians develop enhanced visualization, judgment, and sensitivity through rigorous training and clinical practice, which may explain their higher detection rates. We also hypothesize that clinicians’ superior performance may reflect conscious or unconscious motivational bias (Montibeller & von Winterfeldt, 2015), as they may invest additional effort in identifying asymmetry in regions closely related to their professional expertise.

Interestingly, studies show that laypeople with exposure to a specific condition—for example, individuals who have undergone orthodontic treatment—are better at perceiving asymmetry, such as incisal plane canting or midline shifts, than those without such experience (Naini et al., 2012a; Naini et al., 2012b; An et al., 2014). Similarly, managing cleft-craniofacial conditions poses significant challenges for patients and families, and family members involved in long-term care may develop greater awareness and critical perception of craniofacial symmetry (Denadai & Lo, 2022). This may explain why, in our study, family members demonstrated higher recognition rates of asymmetry compared to laypeople without prior exposure to appearance-altering craniofacial conditions. Our findings suggest that future research should consider subdividing lay observers rather than treating them as a single group. To further explore how technical and experiential backgrounds influence temporal asymmetry assessment, future studies could examine whether other trained healthcare providers (e.g., neurologists, neurosurgeons, psychologists managing epilepsy) and different untrained observers (e.g., family members of patients with or without epilepsy surgery) perceive asymmetry differently. Additionally, research should investigate how various skills and strategies—such as perceptual fluency, cultural influences, artistic experience, focus on specific facial regions, and sensitivity to shape, contour, or light and shadow (Pinheiro et al., 2023)—affect the perception of symmetry versus asymmetry in 3D craniofacial images across different observer groups. However, it should be acknowledged that visually assessing temporal asymmetry, particularly for untrained observers, remains inherently challenging and represents a limitation of such studies.

Similar to many medical and nonmedical contexts (Bathiany, Hidding & Scheffer, 2020; Nakajima, Okuda & Komatsu, 2021), the perception of craniofacial asymmetry appears to follow a threshold model, becoming clinically noticeable only beyond a certain degree of deviation (Parrini et al., 2016; Wang et al., 2017). Studies have examined various craniofacial subunits to define objective cutoff values—using linear, angular, or percentage measures—that correspond to subjective detection by clinicians and laypeople ((Padwa, Kaiser & Kaban, 1997; Parrini et al., 2016; Wang et al., 2017; Lee, Dumrongwongsiri & Lo, 2019)). Each craniofacial subunit has a unique perceptual threshold that distinguishes normal variation from clinically recognizable deviations (Padwa, Kaiser & Kaban, 1997; Parrini et al., 2016; Wang et al., 2017).

To date, no study has quantified the minimal degree of objectively measured temporal asymmetry detectable by both trained and untrained observers. Although a 10% threshold has been proposed for defining visible unilateral temporal deformity (Kim et al., 2018), this value remains unvalidated through formal observer detection analyses. Craniofacial soft tissue thickness—a linear measurement widely used in clinical and research settings for purposes such as evaluating surgery-induced changes, assessing sarcopenia, serving as a prognostic marker in cancer, and aiding facial approximation in forensic medicine (Stephan & Devine, 2009; Hona & Stephan, 2024)—was therefore employed as the objective measure in the current study.

Our cumulative data demonstrated that both clinicians and lay observers were able to identify temporal asymmetry at a threshold of 10% difference in temporal thickness. All observer categories and subcategories met or exceeded the minimum clinical recognition level, with over 90% accuracy in detecting asymmetry in 3D craniofacial images exhibiting temporal thickness differences greater than 10%. This threshold effectively distinguished between 3D craniofacial images predicted to be unaffected (preoperative images) and those hypothesized to exhibit visible temporal asymmetry (postoperative images). Based on our findings, and in alignment with definitions supported by previous research (Kim et al., 2018), a 10% difference in temporal thickness can be proposed as a discriminative threshold for the clinical detection of temporal asymmetry by both trained and untrained observers. Importantly, differences in sensitivity for the detection of temporal asymmetry between clinicians and lay observers could have practical implications. In real-world settings, lower sensitivity among lay observers might mean that minor temporal asymmetries, although noticeable to clinicians, remain imperceptible to the public, potentially reducing patients’ psychosocial distress or perceived need for intervention. Therefore, recognizing this potential clinical scenario could help avoid unnecessary treatment or overinflated diagnoses driven by clinicians’ higher sensitivity. Understanding these dynamics is crucial for guiding patient counseling, setting realistic expectations, and informing decisions about therapeutic interventions.

While our study does not answer all questions about craniofacial asymmetry, it offers valuable insights into how clinicians and lay observers recognize temporal asymmetry. We have also established a key threshold for clinical detection, which may benefit interdisciplinary epilepsy care. This data provides an initial framework for counseling adult patients and setting expectations for potential postsurgical asymmetry. Moreover, it could aid perioperative assessments in temporal reconstruction (Vaca et al., 2017; Gonçalves et al., 2021), where residual asymmetry may exist but remain imperceptible in interpersonal interactions.

Potential caveats of this study should be acknowledged. This was not an epidemiological investigation, and we did not assess the prevalence of temporal deformity, explore multiple levels of nonindependence, or investigate potential predictors of its occurrence following epilepsy surgery. The study was also not designed to examine hierarchical data structures or to model clustered or nested sources of variability. These aspects merit future research using alternative methodologies, including mixed-effects model analyses. Future investigation could explore specific structural components of the temporal region, such as fat, muscle, and bone, to provide a more detailed understanding of asymmetry. Studies could also assess the severity of asymmetry and treatment needs using continuous or ordinal rating scales, as well as additional parameters such as surface area and volume, allowing for more nuanced differentiation beyond binary classification. Additionally, leveraging artificial intelligence or machine learning techniques for automated detection holds promise for matching or even exceeding human performance in identifying temporal asymmetry.

Conclusions

This study identified a 10% difference in temporal thickness as the threshold for the clinical recognition of temporal asymmetry by both trained clinicians (surgeons and clinical specialists) and untrained observers (family members and laypeople). Moreover, significant differences were observed between these groups, with trained clinicians detecting asymmetry more frequently than untrained observers across most conditions—including overall temporal thickness differences, the 10% cutoff, preoperative/postoperative image types, and varying severity levels—except at the highest severity level, where all observers achieved complete detection.

Supplemental Information