Ten human post-mortem brains (five male, mean age 67, range of age 30–85 years) were used for cytoarchitectonical analysis. The brains were obtained from the body donor program of the Department of Anatomy at the University Hospital Düsseldorf of the Heinrich-Heine-University. Ethical approval and documented written informed consent were obtained in accordance with legal requirements (Faculty of Medicine, Heinrich-Heine-University, Düsseldorf, Germany, ethics approval number 4863). The clinical history did not reveal any psychiatric or neurological disease. The histological procedure and subsequent image analysis were performed as described in detail (Amunts et al., 2020). The post-mortem delay was between 12 and 24 hours. After removal of the brain from the skull, the brains were fixed in formalin or Bodian's fixative for approximately 3–6 months, embedded in paraffin and sectioned in 6,000–7,500 coronal slices using a large-scale microtome (20 μm thick serial whole-brain sections). Every 15th section was stained for cell bodies with a modified silver staining method (Merker, 1983). After digitization, every 30th to 60th section (distance between sections: 0.6 and 1.2 mm) was used for observer-independent cytoarchitectonical analysis.

Borders were determined using observer-independent image analysis and statistical tools designed to identify significant changes in the laminar pattern, i.e., the cytoarchitecture (Schleicher et al., 1999, 2009; Amunts et al., 2020) (Figure 2). Images with a resolution of 1 μm/pixel (~8 Gb per image, 8 bit) were obtained from histological sections of 10 human post-mortem brains (Table 1). A high-throughput brightfield microscope (TissueScope LE120, Huron Digital Pathology) was used to scan the sections on which the regions of interest (ROI), including the superior and middle frontal gyrus, were analyzed (Figure 2A).

A Matlab-based script (The MathWorks, Inc., Natick, MA, USA) was used to convert the digitized ROIs into gray level index (GLI) images (Schleicher et al., 2009) (Figure 2B). The GLI represents the cytoarchitectonic organization (Schleicher et al., 1986) by estimating cell body volume fraction (Wree et al., 1982) in a 16 × 16 μm² measuring field. Equidistant traverses were then calculated between the contour lines marking the layer I/II boundary (outer contour line) and the layer VI/white matter boundary (inner contour line) using a physical model of electric field trajectories (Jones et al., 2000). The outer and inner lines were interactively determined in each GLI image using Matlab scripts in-house written. GLI profiles were extracted perpendicular to the cortical layers (Figure 2C). These GLI profiles represent the changes of the GLI from the cortical surface to the white matter and characterize the laminar changes in the cytoarchitecture of the ROI (Schleicher et al., 2009). Each profile was normalized to a cortical depth of 100% in order to compare cortices of different thicknesses. The GLI profiles were characterized by a 10-dimensional feature vector consisting of mean GLI, cortical centroid depth, standard deviation, skewness, kurtosis, and similar parameters for the first derivatives (Schleicher et al., 2009). The advantage of the GLI method is that it is based on the cell packing density and thus minimizes the influence of the color intensities of the brain sections. Although nuclei of glial cells and endothelial cells are also stained, their area-specific areal-specific or laminar differences do not seem to influence the present analysis and parcellation (Wree et al., 1982).

The Mahalanobis distance, a multivariate distance measure, was used to quantify differences between feature vectors of adjacent blocks of GLI profiles (Mahalanobis et al., 1949; Schleicher et al., 1999) (Figures 2D, E). The shapes of adjacent profiles reflected by the ten-dimensional feature vector were compared for observer-independent border detection (Schleicher and Zilles, 1990; Schleicher et al., 2005). Each profile block consisted of 12 to 30 profiles. The distance for each localization along the cortical band was calculated using a sliding window procedure. Maximum Mahalanobis distances indicated areal boundaries (Figure 2F). Hotelling's T2 test with Bonferroni correction for multiple comparisons (p < 0.001) was performed to determine the significance of the boundaries. Each border was verified by visually inspecting the histological images to ensure histological and cutting qualities. Borders were accepted if consistently detected in the same positions across multiple block sizes and at least three adjacent sections.

Hierarchical cluster analysis was performed to detect similarities between the new DLPFC areas and the neighboring frontal pole areas (Fp1, Fp2) (Bludau et al., 2014), anterior DLPFC areas (Bruno et al., 2022), areas 44 and 45 of Broca's region (Amunts et al., 1999), subdivisions of areas 8 [8d1, 8d2, 8v1, and 8v2 (Amunts et al., 2021)] and areas of the motoric region, like 6d1, 6d2, and 6d3 (Amunts et al., 2021), and 4a, and 4p (Geyer et al., 1996). Three sections with an average of 15–20 profiles per hemisphere were extracted for each area and brain, i.e., approximately 45 profiles per hemisphere (900 profiles total). Sections were selected that were free of histological artifacts, large vessels and that were not cut tangentially. Feature vectors were generated for each area, and discriminant analyses were calculated using Euclidean distance and the Ward-linking method (Ward, 1963) based on the mean GLI profiles (Schleicher et al., 1999). A high Euclidean distance indicates high cytoarchitectural difference and low structural similarity, and vice versa. The results were visualized using a dendrogram.

The area extent of both hemispheres was completely labeled in digitized high-resolution histological sections (1,200 dpi; ~20 μm/pixel; 8-bit gray value resolution) by the in-house software “Section Tracer Online Tool”. The same deformation fields calculated for the histological volumes of the 10 post-mortem brains (Amunts et al., 2020) were then used to reconstruct the cytoarchitectonic regions in 3D.

The brain's 3D reconstruction was created using the structural magnetic resonance image (MRI) 3D data set of the fixed brains before sectioning and high-resolution flatbed scans of stained histological sections (Amunts et al., 1999). The datasets were compared, and adjustments were made to account for any deformations or shrinkage (Henn et al., 1997; Amunts et al., 2004; Bludau et al., 2014). To create a stereotaxic map and account for anatomical variability, the delineated areas of all examined brains were normalized and transferred to Montreal Neurological Institute (MNI) Colin27 and the non-linear asymmetric MNI152 2009c (ICBM152casym) template using linear and non-linear elastic registration tools (Evans et al., 1992; Henn et al., 1997; Hömke, 2006). The reference brain templates were used to superimpose the new cytoarchitectonic areas and generate probabilistic maps (Amunts et al., 2020).

These maps illustrate the intersubject variability of a cortical area at a given position in the reference brain. Probabilities were color-coded from dark blue (high variability) to red (low variability). These maps provide a metric of intersubject variability: the higher the probability, the lower the cerebral intersubject variation in that cortical area. These probability maps were used to create a continuous, non-overlapping maximum probability map (MPM) of the newly identified and previously mapped areas, assigning each voxel to the area with the highest probability corresponding to that particular voxel (Eickhoff et al., 2005). Centers of gravity were calculated. All area maps are publicly available via the Julich Brain Cytoarchitectonic Atlas (Amunts et al., 2020) and the HBP Human Brain Atlas as part of EBRAINS (https://ebrains.eu/service/human-brain-atlas).

Interhemispheric and sex differences in the volumes of the new areas were analyzed and compared between brains. An in-house software labeled cortical areas in histological, high-resolution (7,000 × 6,000 pixels), and cortical sections (1.2 mm apart, every 60th section). The volumes of the ROIs were calculated using the following formula:

V is the volume of the ROI (mm²), s the distance between the examined sections (number of sections), T is the thickness of a histological section (20 μm), x is the width of a pixel (0.02116 mm), y is the height of a pixel (0.02116 mm), ΣN_i is the sum of the areas of the structure overall sections (in pixels), and F is the shrinkage factor of each brain. To compare the volumes of DLPFC areas from brains of different volumes, they were individually corrected for shrinkage due to histological processing (Amunts et al., 2007).

Normalized volumes were analyzed using mixed model repeated measures ANOVA (within factor, area and hemisphere; between factor, sex) to test for volume significance. Normality was tested by Shapiro–Wilk and sphericity by Mauchly. Levene's test was used to assess the homogeneity of error variances. A significance level of α = 0.05 was used for all tests.

GLI profiles characterize the changes in the volume fraction of cell bodies (Wree et al., 1982), i.e., the laminar pattern of the areas. An increased volume fraction of cell bodies corresponds to a reduced proportion of neuropil, i.e., less axonal, dendritic, and synaptic space. Mean GLI values were calculated based on 15–20 profiles in three histological sections per area, hemisphere, and brain. Mean GLI values were studied for sex, inter-hemispheric and inter-area differences. Statistical analyses were performed using mixed model repeated measures ANOVA (within factor, area, and hemisphere; between factor, sex) at a significance level of α = 0.05.

All areas SFG2, SFG3, SFG4, MFG4, and MFG5 showed six cortical layers, including layer IV, and, thus, represent a typical granular isocortex. However, individual and neighboring areas could be clearly separated based on their distinct cytoarchitectonic characteristics, such as size, density, and arrangement of neurons within single cortical layers (summarized in Table 2).

Differences in the shrinkage-corrected volumes of the five areas showed that area SFG2 had the largest volume (5,711 ± 324 mm³), followed by SFG4 (4,281 ± 226 mm³), MFG4 (4,031 ± 167 mm³), MFG5 (3,630 ± 112 mm³), and SFG3 (1,504 ± 132 mm³). The combined cortical volume of DLPFC areas was 10,029 ± 239 mm³ in the right hemisphere and 9,128 ± 267 mm³ in the left hemisphere. Male brains had a total volume of 22,981 ± 785 mm³, with 11,648 ± 397 mm³ in the right hemisphere and 11,333 ± 388 mm³ in the left hemisphere. Female brains had a total volume of 15,334 ± 444 mm³, with 8,410 ± 248 mm³ in the right and 6,924 ± 196 mm³ in the left hemisphere.

Differences in the shrinkage-corrected volumes of the five areas were analyzed for interhemispheric and sex differences. Shrinkage-corrected area volumes were normalized to the corresponding total brain volume and then compared using an ANOVA (within factors, area and hemisphere; between factor, sex). No significant volume differences between sexes and hemispheres were found.

The human DLPFC was microstructurally parcellated into distinct areas by former researchers (Brodmann, 1909; Rajkowska and Goldman-Rakic, 1995a; Petrides and Pandya, 1999). However, there is evidence that former maps do not adequately reflect the whole functional diversity of this brain region as described by functional parcellations (Yeo et al., 2011; Shen et al., 2013; Schaefer et al., 2018). As it has already been shown by Bruno et al. (2022) that parts of the anterior DLPFC can be subdivided into four subregions, we now focused on additional adjacent areas using the same methodological approach (Schleicher et al., 1999; Amunts et al., 2020) at the microstructural level. We were able to delineate a further five areas based on their cytoarchitecture.

In contrast to former microstructural maps, the present study provides three-dimensional probability maps in reference spaces with a finer subdivision of the human DLPFC based on reliable observer-independent image analysis. As the cytoarchitectonic parcellation of Petrides and Pandya (1999) is the most recent map of the DLPFC and shows the highest unity with our results concerning subdivisions, we are focusing on that study for comparison. They subdivided the DLPFC into four distinct regions. Area 9 covers the surface of the sfg but also extends medially to the paracingulate sulcus. The main cytoarchitectonic features were compact layers II and IIIa, large pyramidal cells in the deeper layer IIIc and a thin, poorly developed layer IV. Area 46 has a uniform appearance with medium-sized pyramidal cells in layers IIIc and Va and a broad cell-dense layer IV. The area was mainly in the mfs and the middle parts of the mfg. The rest of the mfg was covered by area 9/46, which could be divided into a dorsal and ventral part. Area 9/46 has a well-developed layer II with a distinct boundary to layer III. Furthermore, there are considerably large pyramidal cells in the deeper layer IIIc, a well-developed layer IV, medium-sized cells in layer Va and a sparse layer Vb. In area 9/46d, the pyramidal cells in layer IIIc are less densely packed, and layer IV is thinner than in the ventral part 9/46v.

Compared to the areas delineated in the present study, the cytoarchitectonic descriptions of Petrides and Pandya (1999) fit well, although not entirely. Area SFG2 shows the same characteristics as area 9, but the extent of area SFG2 is limited to the dorsal aspect of the frontal lobe and does not extend to the mesial surface compared to area 9. Area SFG3 has a poorly developed layer IV like area 9 but shows some difference in infragranular layers, such as separable layers V and VI. The microstructural features of 9/46d, like a distinct transition of layer II to III, large pyramidal neurons in deeper layer IIIc, and a well-definable layer IV, can be assigned to the area SFG4. Furthermore, layer V consists of medium-sized pyramidal cells in layer Va and sparse layer Vb. Area 9/46v covers the more ventral parts of the mfg in the former map. However, the newly delineated area SFG4 extends to the sfg and partly to the mfg and mfs and thus does not coincide entirely with the described location of the area 9/46d, which was mainly located on the mfg.

Comparing the area localization of MFG4 on the mfg with the former map of Petrides and Pandya (1999), MFG4 overlaps with the area 9/46v. Both areas have a broader and more pronounced layer IV compared to adjacent areas SFG4 or 9/46d, respectively. However, the boundary between layers II and III was not as distinct, and we could not detect a sparse layer Vb in area MFG4 compared to area 9/46v. Area MFG5 shares some cytoarchitectonic characteristics of the formerly described area 9, like a thin layer IV. However, there were also microstructural differences. For example, layer V was not composed of large pyramidal cells, but neurons in deeper layer IIIc were larger than layer V neurons.

Relating these comparisons with the result of the performed cluster analysis and the previously published data (Bruno et al., 2022), the DLPFC can be described with respect to its cytoarchitectonic organization as follows: Areas MFG1, SFG2, SFG3, SFS2, and MFG5 reveal main characteristics of area 9 as described in former microstructural maps (Brodmann, 1909; Rajkowska and Goldman-Rakic, 1995a; Petrides and Pandya, 1999), like a thin layer IV and large pyramidal cells in the pyramidal layers. However, these areas differ from each other, as shown by image analysis statistical tests, and cannot be merged into a single area. The present parcellation shows a more fine-grained parcellation than previous maps (Brodmann, 1909; Rajkowska and Goldman-Rakic, 1995a; Petrides and Pandya, 1999). Areas SFS1 and MFG2, with their homogeneous appearance and well-developed layers IV, could correspond to area 46 as defined by Rajkowska and Goldman-Rakic (1995a) as well as Petrides and Pandya (1999). Areas SFG4 and MFG4, clustering together, seem to resemble the transition area 9/46 (Rajkowska and Goldman-Rakic, 1995a; Petrides and Pandya, 1999). But they share more cytoarchitectonic similarity with areas MFG1 and SFG2 than to the remaining DLPFC areas (see Figure 6).

The cluster analysis revealed no correlation between cytoarchitectural similarity and topography. In general, DLPFC areas of the sfg clustered with areas on the mfg: For example, SFG2 clustered with MFG1, and SFG4 with MFG4. This result indicates that no clear gradient could be observed along a topographical axis. It is reasonable to assume that the DLPFC has an internal functional organization, and that correlations are more likely to be found on a structural-functional level, as discussed in the following section.

Up to now, nine microstructurally distinct areas within the human DLPFC could be delineated based on the cytoarchitecture, including the newly identified areas and a previous cytoarchitectonic study (Bruno et al., 2022). This new fine-resolved map now allows us to locate functions more precisely from activation studies. A first comparison of the areas with neuroimaging data on brain function suggests distinct contributions that will be discussed with existing functional studies and parcellations (Yeo et al., 2011; Glasser et al., 2016) in the following sections.

Yeo et al. (2011) studied the organization of the human cerebral cortex based on intrinsic functional connectivity, resulting in different functional networks. The DLPFC coordinates of the present study were compared to the study of Yeo et al. (2011) by their center of gravity MNI coordinates (Table 3). This analysis revealed that SFG2 and SFG3 can be associated with the default mode network. SFG4 and MFG4 of both hemispheres could be associated with the salience/ventral attention network, while MFG5 is part of the frontoparietal control network (ControlA).

Several studies have shown that SFG2 is functionally involved in self-reflection, e.g., the area is the location of a common activation at rest and during the self-referential task (D'Argembeau et al., 2005), self-evaluation (Ochsner et al., 2005), self-representation (Debbane et al., 2017), and cognitive empathy (Schnell et al., 2011). The brain's default mode network has become closely associated with self-referential mental activity, particularly in the resting state (Davey et al., 2016). Based on the visual assignment in the functional atlas of Yeo and colleagues, SFG2 can be assigned to the default network (Yeo et al., 2011). When new information is introduced, the ventral attention network shifts the attention (Vossel et al., 2014). SFG2 clusters with MFG1 (Bruno et al., 2022, Figure 6B), which can be assigned to this attentional network.

Little is known about the functional role of the region where area SFG3 is located. One imaging study has shown that the area is active in fragment-based retrieval in the right DLPFC and thus plays a role in episodic retrieval (Allan et al., 2000). The area has also been shown to be active in episodic memory (Smith et al., 2007). Episodic memory retrieval processes with strong memory content recruit the default network (Kim, 2016), and the successful retrieval from episodic memory involves the frontoparietal network (Iidaka et al., 2006). Thus, the results of the imaging studies are consistent with the network analysis (Yeo et al., 2011), that SFG3 could be assigned to the default network. SFG3 clusters with MFG5, an area that could be allocated to the frontoparietal control network (ControlA, Figure 6).

The coordinates for SFG4 seem to be implicated in the salience/ventral attention network (Yeo et al., 2011). The salience network is essential in tasks involving attention and responding to unexpected but salient stimuli (Sridharan et al., 2008; Menon and Uddin, 2010). The bottom-up attentional process involves the ventral attentional network (Vossel et al., 2014). Imaging studies have shown activation of the area in error awareness (Masina et al., 2018) and self-control (Hayashi et al., 2013). SFG4 clusters with MFG4 according to cytoarchitectonic similarities. However, like SFG4, the coordinates of the probabilistic map of MFG4 seem to be part of the salience/ventral attention network (Yeo et al., 2011). A previous study attributed MFG4 to working memory connected to the anterior cingulate cortex (Cieslik et al., 2013). Further studies found activations in the center of MFG4 during successful self-control (Harris et al., 2013; Chen et al., 2018; Ren et al., 2023) and self-regulation leading to better working memory performance (Zhang et al., 2013). Several studies consider the DLPFC as part of a large attentional network involving perceptual, attentional, and working memory (Yamasaki et al., 2002; Buschman and Miller, 2007; Zanto and Gazzaley, 2009; Lennert and Martinez-Trujillo, 2011). This broad functional profile of MFG4 can be summarized as attentional filtering to support self-control by reducing irrelevant stimuli (Zanto and Gazzaley, 2009; Harris et al., 2013).

MFG5 could be bilaterally assigned to the frontoparietal control network (Yeo et al., 2011), which is supported by other comparisons with functional activation studies. For example, DLPFC-mediated top-down control interferes with contextual memory-guided attention (Rosero Pahi et al., 2020), and MFG5 is active in the self-control of choice behavior (Schmidt et al., 2018). The frontoparietal network of the DLPFC is involved in cognitive control and working memory (Harding et al., 2015; Marek and Dosenbach, 2018). Indeed, MFG5 has also been shown to be involved in working memory processes, particularly in processing verbal input (Henson et al., 2000; Barbey et al., 2013).

In the multimodal parcellation of Glasser et al. (2016), the DLPFC is subdivided into 13 areas, including areas from the posterior part (8Ad, 8BL, 8Av and 8C) and premotor region (s6-8, i6-8, and SFL). Based on the results of the cluster analysis in this study (see Figure 6), which shows a clear separation of the posterior areas from the DLPFC areas, we suggest, like others (Jung et al., 2022), that these should not be included in the DLPFC by definition. Without these posterior areas, the DLPFC is comprised of the six areas 9a, 9p, 9-46d, 46, a9-46v, and p9-46v in the Glasser atlas (Glasser et al., 2016). Results from this and a previous study (Bruno et al., 2022) lead to a parcellation with so far nine areas (SFG2, SFG3, SFG4, SFS1, SFS2, MFG1, MFG2, MFG4, and MFG5). Based on a currently ongoing mapping study, we assume that the DLPFC can be microstructurally divided into at least 10 distinct areas.

Comparing the Glasser et al. (2016) map with our parcellation (see Figure 8), similarities as well as discrepancies can be found and will be discussed in the following. With its position on the most dorsal part of the sfg and borders to a subdivision of area 8, the area SFG2 matches the described area 9p. The same conditions are present for the areas SFG3 and 9a, which are adjacent to the frontal pole area (Fp1 or 10p) and located on the more anterior parts of the sfg. However, regarding the probability maps of our areas, area SFG3 is located to a large extent within the sfs and not on the surface of the sfg.

Area SFG4 shows similarities with the described area 9-46d, as it shares borders with the areas on the sfg (SFG2 and SFG3 or 9a and 9p, respectively). Regarding the entire area extension and the posterior borders, the areas SFG4 and 9-46d do not correspond. Area SFG4 borders posteriorly to the subdivisions of area 8, like area 8d2 and 8v2, whereas 9-46d mainly borders area 46 and only to a very small extent with area 8Ad (Glasser et al., 2016). This difference that area 46 borders posteriorly to areas 8Ad and 8Av (Glasser et al., 2016) is not given in our current map and cannot be found in any former cytoarchitectonic maps (Brodmann, 1909; von Economo and Koskinas, 1925; Sarkissov et al., 1955; Rajkowska and Goldman-Rakic, 1995b; Petrides and Pandya, 1999).

Area MFG4 can be related to the described area p9-46v (Glasser et al., 2016) concerning its location on the mfg dorsally to the ifs areas and anteriorly to the subdivisions of area 8. In contrast, the total area volume of MFG4 seems considerably larger compared to area p9-46v.

Area MFG5 shows similarities with the described area IFSa in its shape and location dorsal to Broca's area 45. However, area MFG5 was located to notable parts on the surface of the mfg and only to some extent in the ifs compared to the exclusively sulcal area IFSa (Glasser et al., 2016). Therefore, and in regard to the results of the cluster analysis (see Figure 6), we prefer the definition as MFG area and the assignment of the DLPFC instead of the VLPFC.