There is a moment in a hydrocephalus clinic that no amount of software has yet dissolved. The scan is on the screen, the ventricles are plainly large, the walking is plainly wrong, and the family asks the only question that matters: will the operation help him? The honest answer is that nobody in the room knows. Shunt surgery for idiopathic normal pressure hydrocephalus (iNPH) is one of the few operations in dementia care that can give function back, and it is also an operation a substantial minority of people undergo without lasting benefit. So the field keeps returning to the picture, hoping the geometry of the fluid spaces holds the answer.
A study published in Fluids and Barriers of the CNS on 9 April 2026 measured that geometry more carefully than most. Andrius Penkauskas and colleagues at the University of Eastern Finland took 170 patients from the Kuopio NPH registry, 95 of whom were classed as responders and 75 as non-responders using symptom assessments before surgery and one year after it. Preoperative T1 MRI scans were segmented with cNeuro software, three dimensional meshes of the lateral ventricles were built, and 27 geometric features were extracted: global shape, local shape, curvature. A regularised machine learning model called LogitNet was trained and tested with repeated nested cross validation, ten rounds, five inner folds for tuning and five outer folds for evaluation. This is retrospective work on one registry's scans, and it was designed as a proof of method rather than a clinical tool.
The result is the interesting part, because the authors report it straight. The model reached an area under the curve of 0.661, with a standard deviation of 0.066. AUC runs from 0.5, which is a coin toss, to 1.0, which is perfection; 0.661 is a coin weighted slightly in your favour. A single feature carried almost all of it: asphericity, essentially how far the ventricle departs from a sphere. Asphericity on its own scored 0.660, meaning the other 26 measurements added next to nothing. Reducing the features by principal component analysis made things worse, dropping the AUC to 0.604. In their discussion the authors place this against earlier work using the callosal angle to predict shunt outcome, citing an AUC of 0.640 from Grahnke and colleagues in 2018 and 0.620 from Gao and colleagues in 2025, while noting that their own figure falls short of the 0.720 that Gao's team reached by combining DESH, the callosal angle and the tap test. The lead author's own framing is modest: the 3D geometry, he said, can "provide important clues about which patients are likely to benefit from surgery". Clues, not answers. There is no external validation cohort, and two of the co-authors hold shares in the parent company of the firm behind the segmentation software, a competing interest the paper declares.
Read this next to a review published in Biomedicines on 16 March 2026, in which Juan Sahuquillo, Maria Antonia Poca and colleagues argue for redefining iNPH using AI-driven brain volumetry. Their case is that the Evans Index, the callosal angle and the DESH pattern are two dimensional proxies with real interobserver variability, and that automated volumetry of ventricular and extraventricular cerebrospinal fluid, grey matter, deep grey nuclei and periventricular white matter could describe the condition as what it probably is: a continuum running from predominantly hydrocephalic to mixed states where impaired CSF dynamics and neurodegeneration sit in the same brain. That is a serious and plausible argument about classification. It is not the same argument as predicting who benefits from an operation, and the two get conflated constantly in coverage of NPH imaging. Describing a syndrome better and forecasting a surgical outcome are separate claims requiring separate evidence, and the Finnish numbers are a reminder of how much harder the second one is.
For a family, the practical shape of this is unchanged, and worth saying plainly. Selection for shunt surgery today still rests on clinical assessment, the pattern of symptoms, and CSF drainage testing, judged by a specialist team, not on a number generated from the shape of a ventricle. The PENS trial, a double blind randomised placebo-controlled study published in the New England Journal of Medicine online on 16 September 2025, enrolled participants whose gait velocity improved after temporary CSF drainage and found that an open shunt valve produced significant improvement at three months in gait velocity and in a measure of gait and balance, but not in the cognitive or bladder measures used. That is meaningful evidence that shunting does something real for carefully chosen people, and it also shows how narrow and how short the demonstrated benefit currently is. An imaging model scoring 0.661 cannot refine that choice, and would misclassify too many people to justify letting it try.