Cognitive models of lexical retrieval have been heavily influenced by visual confrontation (picture naming) tasks. Further, existing studies have implemented methods that either lack causal certainty or precise spatial localization. Therefore, published results are mixed with respect to whether visual and auditory stimuli represent overlapping (convergent) or non-overlapping (divergent) neural systems. This unresolved conflict has direct implications on cognitive rehabilitation strategies for brain tumor patients with dysnomic aphasia. In this study, we propose a convergent model of lexical retrieval within a lesion-symptom framework for both visual and auditory inputs. First, we established that intrinsic brain tumors cause extensive disruption of normal cytoarchitecture (Fig. 1), leading to selective impairments in word finding (Fig. 2). This finding is in line with previous reports of the prevalence of selective dysnomia in patients with dominant hemisphere intrinsic brain tumors56. Furthermore, it supports the use of our brain tumor lesion model in particular to study lexical retrieval, since participants in other clinical populations tend to have additional confounding language impairments57,58.
Next, we demonstrated strong correlations between accuracy on picture naming and auditory naming, providing initial evidence of the existence of lexical retrieval pathways that are agnostic to the sensory modality of the input (Fig. 3). From a behavioral perspective, this finding replicates those in lesion studies performed in several clinical populations. For instance, Hamberger and Seidel (2003) found that patients with left temporal lobe epilepsy had co-occurring impairments in visual and auditory naming compared to a) healthy controls and b) patients with right temporal lobe epilepsy59. Miller et al. (2010) and Hirsch et al. (2016) also reported significant associations between picture naming and auditory naming in lesion studies of patients with dementia60,61. Hirsch et al. in particular used PCA in a cohort of 458 patients to reveal a unique redundancy between picture naming and auditory naming that was not shared with any of their other twenty-five cognitive and linguistic measures. In the present study, PCA led to analogous results: the similarity in loadings between PN, AN, and word finding argue for a single linguistic construct (i.e., lexical retrieval) that may be differentiated from constructs contributing to other linguistic measures such as TR and Syn.
As predicted by the results of our language assessments, multivariate lesion-symptom mapping revealed overlapping clusters for picture naming, auditory naming, and PC1 in the lateral prefrontal cortex (Fig. 5A-C). Notably, in contrast to prior DES studies12,62, a separate cluster specific to auditory naming was not identified in the anterior temporal lobe, a well-represented region in our lesion overlap mask. Our results instead align with a number of more recent, mixed-method studies that argue for a convergence hub in the lateral PFC for both visual and auditory inputs14,63. Specifically, by timing the evolution of cortical responses to the onset of task stimuli using electrocorticography, Forseth et al. (2018) showed that the inferior frontal gyrus (a component of the lateral PFC) serves as an interface between lexical and phonological pathways during both picture and auditory naming14. Taken with the results of the present lesion study, these findings provide compelling and causal evidence that lexical retrieval indeed represents a unified cognitive construct that is agnostic to the input modality.
Notably, case reports and small clinical series of patients with modality-specific naming impairments such as visual (i.e. optic aphasia), auditory, and tactical anomia from brain lesions may seem to argue against the existence of a single construct subserving lexical retrieval64–66. However, these studies have not been widely validated in large clinical cohorts using rigorous image-based methods and thus cannot be adequately evaluated for confounding variables. Specifically, concurrent damage to primary sensory processing or association areas cannot be ruled out without the granularity provided by voxel-based analyses. For instance, Hamberger and Seidel (2009) reported that a cohort of fourteen patients with anterior temporal lobe lesions (defined as < 5 cm from the temporal pole) had impaired auditory, but not picture naming16. However, of those fourteen patients, only five had structural lesions. The remaining “anterior lesion” patients were defined by the presence of seizure foci on subdural or scalp EEG. Given a) the absence of lesion overlap maps for examination and b) the ability of epileptic foci to transiently impair remote brain areas, we cannot rule out the possibility that the observed differences in auditory and picture naming performance were due to confounding damage to auditory sensory processing areas (i.e. A1-A3). Indeed, the “auditory-only” naming sites identified by Malow et al. (1996) via DES (i.e. transient induction of iatrogenic lesions) are chiefly located in the superior temporal gyrus where various features of auditory stimuli are encoded67–69. Analogous concerns have also been raised about other modality-specific anomia syndromes70.
Since the late 19th century, it has been known that the brain, at its most fundamental level, is made of individual cells which together form discrete networks that influence cognition and behavior. Efforts to uncover the role of neuroanatomic structures and underlying network dynamics have established both causal and correlative cognitive frameworks. However, these physiological models may lack disease-specific relevance. Therefore, the discovery of overlapping neural systems for lexical retrieval in a brain tumor lesion-symptom mapping study has implications beyond causal adjudication between established scientific models. One salient example comes from competing rehabilitative strategies for patients with aphasia. On the one hand, compensatory training relies on the use of alternative stimuli through strategies such as paced speech, associative cuing, and verbal circumlocution to leverage distant yet overlapping neural systems for a given sensory domain. Compensatory training may therefore be considered most promising in the setting of divergent visual and auditory lexical retrieval systems18–23. If, however, distinct sensory stimuli converge on a single brain region and/or network, it may be best if cognitive rehabilitation strategies focus on environmental interventions, such as supportive communication strategies following a lesion to that region24,25. The optimal rehabilitation strategy for patients with brain tumors and subsequent aphasia continues to be an area of active investigation. Therefore, causal evidence using disease-specific models of physiology will contribute to a nuanced understanding of therapeutic options.
Limitations of the Present Study
It is first worth discussing the typical limitations that apply to any lesion-symptom analysis, namely that conclusions about behavioral contributions of brain areas lying outside our lesion masks cannot be made. For this reason, we expanded our analysis to include voxels in which at least two patients overlapped, rather than more restrictive cutoffs of five or more found in previous studies. By performing multivariate comparisons on a high-performance computer that can support these increasingly memory- and computationally intensive analyses, we were able to avoid the reductions in statistical power that typically constrain studies employing mass-univariate tests with post-hoc corrections for multiple comparisons. While lowering the lesion overlap threshold may increase exposure to outlier effects, our implementation of non-parametric statistics for lesion-symptom mapping mitigate this concern by calculating exact p-values without any a priori assumptions about the underlying distribution. In our study, as in others, the impact of outliers was greatly attenuated by performing 10,000 random permutations of the input behavioral data to generate a null distribution for each test statistic73. Despite these measures, however, our lesion mask was not able to capture posteroinferior regions of the perisylvian network, thereby restricting our conclusions to the frontal cortex, anterior temporal cortex, and angular gyrus.
Furthermore, while there is some debate on whether brain tumors can serve as reliable clinical models for lesion-symptom analyses, this position can be extended to stroke for which VLSM was developed and is most commonly implemented. Furthermore, brain tumor histology confirmed disordered neural structures, increased cellularity, and altered cytoarchitecture (Fig. 1). Because perfect clinical models for brain lesions likely do not exist, the literature may be best served by an increase, rather than a decrease, in the diversity of the clinical populations under study74.