The scan a family sees on a clinic screen is usually explained in terms of the cortex: the wrinkled outer sheet, thinned at the temples, thinned at the front. What relatives often notice first, though, is nothing so tidy. It is a flattening of interest, a person who no longer reaches for the things they used to reach for, or a name that will not come. A study published online on 29 October 2025 in the Journal of Alzheimer's Disease goes looking beneath the cortex, at structures that rarely get named in a radiology report: the mediodorsal, anteroventral and pulvinar nuclei of the thalamus, the nucleus accumbens deep in the reward circuitry, and the claustrum, a thin sheet of grey matter wired to much of the cortex and still argued over by neuroscientists.
The reason nobody named them before is largely technical. Arit Banerjee and Manojkumar Saranathan, of the Laboratory for Neuro-Analysis and Imaging at UMass Chan Medical School, work with a pipeline called sTHOMAS, which takes an ordinary T1-weighted MRI, synthesises the white-matter-nulled contrast that would normally require a special scan, and then parcels out thalamic nuclei and neighbouring deep grey structures. The method itself was published on 23 September 2025 in Human Brain Mapping. In the new paper the team ran it retrospectively over scans already collected by two research databases: 380 participants from ADNI, spanning cognitively normal, mild cognitive impairment and Alzheimer's disease, and 274 from the Frontotemporal Lobar Degeneration Neuroimaging Initiative, covering behavioural variant FTD, semantic variant FTD and progressive non-fluent aphasia. Of those, 237 had a second scan, allowing change over time to be measured. Diagnoses were clinical, made under established criteria. Group differences were adjusted for age and sex, effect sizes reported as Cohen's d, and cognitive links tested by partial correlation.
What came out is a divergence of degree and pattern rather than a clean split. Both diseases showed pronounced atrophy in the mediodorsal, anteroventral and pulvinar thalamic nuclei, the nucleus accumbens and the claustrum. In behavioural variant FTD the loss was significantly greater, with effect sizes against cognitively normal participants that are unusually large for structural imaging: the authors report d = 2.043 for the mediodorsal nucleus and d = 1.84 for the anteroventral, both at p < 0.001. In the Alzheimer's arm the picture was more graded, with the same limbic thalamic nuclei showing emerging change at the mild cognitive impairment stage and widening in the Alzheimer's group, alongside reductions in the accumbens and claustrum. Over the follow-up scans, progression was fastest in behavioural variant FTD. Volumes in these regions tracked with executive, language and semantic test scores. That is a correlation measured at one or two points in time; it does not establish that the shrinkage produced the deficits.
The part that will draw the most attention, and deserves the most caution, is the classification. A hierarchical random forest model trained on these volumes separated Alzheimer's from FTD with 96.8% accuracy and an AUC of 0.99, and told progressive non-fluent aphasia from semantic variant FTD with 77.7% accuracy and an AUC of 0.83. Read the numbers in the order the model works, though, and the modesty shows: the first stage, simply separating cognitively normal from dementia, managed 76.3% accuracy with an AUC of 0.85. In other words, the model was best at a question the clinic has usually already answered, and weakest at the one families actually arrive with. And it was trained and tested inside the same two research cohorts, which are volunteers who consented to years of scanning, not the general run of people walking into a memory clinic.
So what does this change for someone living with a diagnosis this week? Nothing, directly. Nobody can order a claustrum volume and get a diagnosis from it, and this study does not tell any clinician to alter a single decision. What it does is add a comparative structural map where there was mostly a gap, in a distinction that matters enormously in practice: Alzheimer's and FTD are still confused with one another for years at a time, at real cost to families who are told the wrong thing about what is coming. The authors are open that prediction is an ambition rather than a result; Banerjee has said the work "lays the groundwork for true future prediction". That groundwork is what has been published. The prediction has not.