https://atlas.rsna.org/schemas/2025-11/model.json

Brainstem and Ventricular MR Planimetry in Patients with Progressive Supranuclear Palsy

https://dx.doi.org/10.1148/ryai.230151

1.0

S.N.; email: moc.liamg@orgin.olegnaerotavlas

Supported by Regione Puglia and CNR for Tecnopolo per la Medicina di Precisione, DGR number 2117 of November 21, 2018 (CUPB84I18000540002), and the Research Center of Excellence for Neurodegenerative Diseases and Brain Aging (CIREMIC), University of Bari, Aldo Moro. Data used from PPMI, 4RTNI, FTLDNI, and ADNI consortia (see Acknowledgments).

Retrospective study; all individuals provided informed consent; protocol approved by the institutional review board at all sites.

Encoder-decoder convolutional neural network: ResNet-50 encoder pretrained on ImageNet + U-Net decoder (ResNet-50–U-Net). Implemented in Keras (Python 3.6.9, TensorFlow 2.4.1). Trained 5 epochs with Adam (initial learning rate 0.01) and data augmentation (rotations −30° to 30°, translations −20% to 20%).

Pretrained ResNet-50 weights: https://github.com/fchollet/deep-learning-models/releases/download/v0.2/resnet50_weights_tf_dim_ordering_tf_kernels.h5; Implementation framework: https://github.com/keras-team; Segmentation codebase referenced: https://github.com/divamgupta/image-segmentation-keras

Automates segmentation and planimetric measurements of brainstem and ventricular structures to support diagnosis of progressive supranuclear palsy and differentiation from Parkinson disease.

Clinical decision support systems; workflow optimization (fast automated measurements).

Reported optimal classification cutoffs (automated): MP = 0.20; MRPI = 12.94; MRPI 2.0 = 2.29.

Fully automated segmentation and measurement pipeline (no user interaction reported).

Automated computation of brainstem and ventricular planimetric measures (midbrain and pons areas; widths of MCP, SCP, third ventricle, and frontal horns) from T1-weighted brain MRI to assist in differentiating progressive supranuclear palsy from Parkinson disease in adults in neuroradiology settings.

Preprocessed volumetric T1-weighted brain MR images (1.5T/3T; multi-vendor); preprocessing with ART to reorient, define midsagittal section and AC/PC; resampled to 256×256×512-mm3; min–max intensity normalization (0–1) on 2D sections.

1) Preprocess T1-weighted MRI with Automatic Registration Toolbox (ART) to reorient, detect midsagittal plane and AC/PC; resample to 256×256×512-mm3. 2) Apply min–max intensity normalization to 2D sections. 3) Run the ResNet-50–U-Net models to segment midbrain, pons, MCP, SCP, third ventricle, and frontal horns. 4) Compute planimetric measures: midbrain/pons areas (pixel counts); MCP width (distance between upper/lower profiles on contour); SCP width (max distance between medial/lateral borders across three sections, averaged); third ventricle width (max lateral-lateral distance); FH width (max left-right distance across axial sections). 5) Derive MP, MRPI, MRPI 2.0 per published formulas.

Manual ground truth by a single expert rater; PD and PSP not pathologically confirmed; only PSP–Richardson syndrome evaluated; no early stopping or cross-validation during training (potential overfitting); min–max intensity normalization may be influenced by outliers; 3% failure rate due to incorrect midsagittal section identification; indices are derived from correlated morphologic features (collinearity).

Use to rapidly obtain planimetric brainstem and ventricular measurements and PSP imaging biomarkers to support radiologist assessment.

Models trained on controls and tested on internal, external, and clinical datasets from multiple vendors and field strengths; consistent Dice >0.85 across datasets; similar AUCs to manual measurements; 97% automated completion rate.

Approximate total computational time <2 minutes per case on NVIDIA Quadro RTX 4000 (8 GB GPU).