They report an AUC of 1.00 for distinguishing DESH-positive iNPH from healthy controls and other neurodegenerative dementias. That sounds suspect—let us see why it is not.
Idiopathic normal pressure hydrocephalus (iNPH) causes cerebrospinal fluid (CSF) to build up in the brain's ventricles. This leads to a reversible form of dementia. It is characterized by gait instability, urinary incontinence, and cognitive decline. Currently, diagnosing iNPH is a gamble. Doctors rely on non-specific clinical signs or invasive "tap tests." These are high-volume lumbar punctures that drain CSF to see if symptoms improve. Such tests are risky for frail elderly patients and often produce inconsistent results.
Radiologists look for "DESH" (disproportionately enlarged subarachnoid-space hydrocephalus). This is a pattern where the Sylvian fissures (deep grooves separating the frontal and parietal lobes) are enlarged. Meanwhile, the sulci (grooves on the brain surface) at the top of the brain remain tight. Spotting DESH is currently a qualitative exercise. It relies on a human eye estimating ratios. This makes it subjective and prone to inter-observer variance.
The Problem
The status quo relies on linear markers like the Evans Index. This is a simple ratio of ventricular width to skull diameter. It only measures how big the holes are. It does not show how fluid is redistributed. It fails to distinguish iNPH from general brain atrophy seen in Alzheimer’s or vascular dementia. In those cases, ventricles also enlarge as brain tissue dies. Even the Callosal Angle, an angular measurement of the corpus callosum, lacks granularity. It cannot easily track disease progression or treatment response. As shown in, a massive gap exists between the signatures of NPH with DESH and other dementias.
Current linear tools cannot resolve this.
How It Works
The authors move from linear measurements to a volumetric ratio called the Disproportionately Elevated Sulcal Index (DESI). They use a 2D-to-3D pipeline. Instead of a heavy 3D U-Net, they segment 2D coronal slices and reconstruct the volume.
The process follows these steps: 1. Alignment: Using BrainSignsNet, the system localizes the anterior and posterior commissures (AC-PC line). This standardizes the brain's orientation. 2. Segmentation: A U-Net with a pretrained EfficientNet-B0 encoder processes coronal slices. It uses a composite loss function. This combines the Dice similarity coefficient (overlap measure) and Hausdorff distance (boundary error penalty). This delineates the Sylvian fissures and superior sulcal spaces. 3. Volumetric Integration: The system sums voxels from these 2D masks. To normalize for brain size, the superior sulci volume is restricted. It uses a 3D wedge extending 30° superiorly from the AC-PC line .
- Indexing: $\text{DESI} = \frac{\text{Volume}{\text{Sylvian Fissure}}}{\text{Volume}$.}}
Numbers
The compute footprint is modest. The model trained on a workstation with dual NVIDIA RTX 4090 GPUs and 128 GB of RAM. The U-Net achieved Dice scores of $\sim 0.80$ for Sylvian fissures and $\sim 0.74$ for superior sulci. These scores indicate a high degree of spatial overlap with manual labels.
The classification results are the headline. At a threshold of 4.83, the paper reports perfect classification (AUC = 1.00). This separates DESH-positive NPH from healthy controls, Alzheimer's, and vascular dementia. For the harder task of separating NPH with DESH from NPH without DESH, the AUC is 0.991. Accuracy reached 97.9% at a threshold of 6.10. Crucially, shows DESI has only a modest correlation with age ($r = 0.178$).
This suggests the index tracks pathology rather than normal aging.
What's Missing
Three main gaps remain. First, the "ground truth" was determined by a single board-certified neuroradiologist. There were no inter-rater reliability checks. We do not know if the AUC reflects AI brilliance or just one person's interpretation.
Second, the diagnostic groups came from different cohorts (BLSA vs. PENS). The authors used N4 bias correction (intensity inhomogeneity reduction) and normalization. Still, they admit residual batch effects from different scanners might have aided group separation.
Finally, the study is cross-sectional. The authors claim DESI is ideal for longitudinal monitoring. However, they have not shown it working on a single patient over time. We do not know the "drift" of this index during shunt response.
Should You Prototype This
Yes, if you build clinical decision support for neurology. Moving from a qualitative "looks enlarged" assessment to a volumetric ratio improves reproducibility. Hardware requirements are low. This could run as a sidecar service in a PACS (Picture Archiving and Communication System) environment. It would not require a GPU cluster. Code is available on GitHub (SiavashShirzad/Disproportionately-Elevated-Sulcal-Index-DESI-). You must request pretrained weights from the authors. Given the high AUCs and the geometric logic of the 30° wedge, this is a viable pilot candidate.