Faster, More Practical, but Still Accurate: Deep Learning for Diagnosis of Progressive Supranuclear Palsy

See also the article by Nigro et al in this issue.

Bahram Mohajer, MD, MPH, is an incoming radiology resident at the University of Pennsylvania. He is a member of the Radiology: Artificial Intelligence trainee editorial board and the Radiological Society of North America Artificial Intelligence (AI) Committee. Prior to his residency, he completed a research fellowship at Johns Hopkins University, during which he received several awards, including the Radiology Alliance for Health Services Research Harvey L. Neiman Award from the Association of University Radiologists. His research focuses on investigating sources of bias in AI models.

Progressive supranuclear palsy (PSP) is a neurodegenerative movement disorder that manifests clinically with axial rigidity, pseudobulbar palsy, supranuclear ophthalmoplegia, and prominent postural instability. Due to its overlapping clinical manifestations with Parkinson disease (PD), PSP frequently is misdiagnosed, especially in its early stages (1). Differentiating these disorders is crucial for tailoring appropriate treatment strategies.

Since PSP’s original description in 1962, many attempts have been made to identify reliable distinguishing features from PD. The guidelines for the diagnosis of PSP have been updated recently in an attempt to broaden the spectrum of patients in this category and to add more sensitive biomarkers (2). The research forefront is exploring clinical examination, behavioral assessment, genetic testing, and, more recently, imaging biomarkers (3). It has been shown that midbrain atrophy, the “hummingbird” sign in the sagittal view of the brain, may be characteristic of PSP. Attempts have been made to propose more accurate volumetric or planimetric compound measures to distinguish PSP from PD, yet their full potential is hampered by the complexity and time-consuming nature of manual image segmentation.

In this issue of Radiology: Artificial Intelligence, a study by Nigro and coauthors (4) aimed to tackle the time-consuming and computationally heavy nature of calculating these imaging biomarkers, using deep learning (DL) methods. This study can be counted as the latest advance in Nigro and colleagues’ 10-year search for clinically applicable brain atrophy indexes for the detection and discrimination of parkinsonism syndromes. These indexes include the atrophy of brainstem anatomic components and compound measures such as the midbrain to pons ratio (MP), the Magnetic Resonance Parkinsonism Index (MRPI), and MRPI 2.0 (5). MRPI is calculated by multiplying the pons area to midbrain area ratio by the middle cerebellar peduncle width to superior cerebellar peduncle width ratio. It is shown to differentiate patients with PSP from those with PD, with an accuracy of more than 90% (5). Nigro and colleagues developed automated, but not DL-based, preprocessing and calculation methods for these measures in 2020 (https://mrpi.unicz.it/).

In their current study, Nigro et al developed a fully automated, DL-based method for MRI planimetric segmentation of brainstem and ventricular structures. They trained their U-Net–based segmentation model on manual segmentation of T1-weighted brain MR images in 84 healthy control participants of the Parkinson’s Progression Markers Initiative study. They tested segmentation performance on one internal and two external sets, each including 35 healthy control participants from other publicly available datasets. They further used these measures and calculated MP and MRPI to differentiate patients with PSP from those with PD in a clinical test dataset. The segmentation accuracy was high, with mean Dice coefficients exceeding 0.85 for all regions across test sets, indicating strong agreement with manual segmentations by a sole radiologist with 30 years of experience. The authors also showed that with the results of their DL model segmentations, they can accurately calculate MP, MRPI, and MPRI 2.0. The automated measurements achieved remarkable diagnostic performance, with an area under the receiver operating characteristic curve above 0.92 for distinguishing between PSP and PD, mirroring the accuracy of manual segmentation methods.

In my opinion, one of the main values of this study lies in making brainstem planimetric measures more applicable and available for routine clinical practice. Nigro and colleagues’ DL-based segmentation model can substantially reduce the time and effort needed for planimetric segmentation and the calculation of brain atrophy indexes. Compared with manual or conventional automated (eg, atlas-based) segmentation methods, the use of this DL-based segmentation is comparably accurate while being faster. Moreover, compared with conventional automated methods, it has simpler preprocessing, needs less computing power, has fewer occasions where the model fails, and when it does, it is easier to fix (eg, identification of the midsagittal section). These features ease the application of the model for early detection in patients with PSP, who can benefit tremendously from having these specific measures in clinical practice.

Among previous studies, Sun et al (6) have used DL-based pipelines for brainstem segmentation and MRPI calculation. Their model had significantly lower absolute percentage errors in MRPI calculation compared with conventional methods of MRPI calculation (5). However, segmentation performance was reported only for the whole brainstem (planimetric segmentation Dice coefficient: 0.91), which is not comparable to the present study by Nigro et al. It is imaginable that the Dice score of smaller and harder-to-segment brainstem structures, namely the superior cerebral peduncle, is possibly significantly lower. Moreover, Nigro et al did not report errors in MRPI calculation, making it similarly incomparable between studies. More recently, Chougar et al (7), using a 2-year retrospective design, assessed the performance of a logistic regression model trained on brain MRI volumetric data (and not planimetric data) for the early diagnosis of parkinsonism and differentiation of PD from PSP. The logistic regression model, while having higher diagnostic accuracy than radiologists blinded to clinical data for the detection of uncertain PD cases (accuracy 89.6% vs 67.9%), performed considerably worse than radiologists in differentiating patients with PSP (accuracy of 62.5% vs 87.1%). Collectively, the results of these studies, along with Nigro et al’s (4), reinforce the use of DL-derived planimetric measurements of the brainstem, such as MP and MRPI, over more comprehensive volumetric measures of brain cortex and brainstem, when balancing practicality and accuracy.

Nigro et al’s study may have several limitations worth mentioning, including the ones the authors mentioned in their Discussion. The main limitation of this study, as the authors also mention, may be the inclusion of only patients with PSP–Richardson syndrome, which is the most common subtype of PSP (8). Studies have shown that different planimetric measures have varying accuracies for PSP subtypes. The authors also discuss collinearity among the calculated imaging biomarkers of MP, MRPI, and MRPI 2.0, considering that these are computed based on similar brainstem and ventricular morphologic features. Also, while not a limitation, it is worthwhile to explore how a DL classifier trained on segmentation model results (or even their radiomics features) performs compared with MRPI or MP when testing the accuracy of diagnosing PSP versus PD. Indeed, there might be a great opportunity to generate more specific and sensitive biomarkers that may not be readily visible to radiologists’ eyes. Finally, some studies have shown that adding brain cortex volumetric or surface area data to a DL-based classifier may increase the accuracy of diagnosing PSP subtypes (9). However, one needs to balance other aspects, like computing demand and preprocessing steps, before forming a judgment about clinical implementation.

In conclusion, Nigro et al’s study showed that DL-based segmentation models are as good as manual segmentation for planimetric measurements for the diagnosis of PSP and can save time and computational processing. Prospective studies are needed to show the real-world clinical implications of such models before clinical implementation. Moreover, future studies with the addition of multimodal imaging, including novel τ PET tracers (10), may further increase the accuracy of early PSP detection.