MRI-to-CT Synthesis With Cranial Suture Segmentations Using A Variational Autoencoder Framework

TL;DR

This work addresses the need for radiation-free pediatric cranial assessment by converting routine T1-weighted MRIs into CT-like images that preserve skull morphology and enable suture analysis. It introduces a dual-stage variational autoencoder framework: first, MRI-to-CT synthesis (sCT) using a MAISI-based VAE with adversarial and reconstruction losses; second, atlas-guided segmentation that combines the sCT with an anatomical atlas to produce probabilistic heatmaps and eight-label skull segmentation (bones plus sutures). The method achieves high-fidelity sCTs (FID around $1.01$) and strong segmentation performance (bone DSC ~ $0.88$, suture DSC ~ $0.79$), with statistical equivalence to real CT in most regions (TOST, $p<0.05$) and significant improvements over baselines ($p<0.05$). This enables radiation-free, suture-aware cranial evaluation in early childhood and supports longitudinal and population studies, while noting limitations related to single-center data and age range.

Abstract

Quantifying normative pediatric cranial development and suture ossification is crucial for diagnosing and treating growth-related cephalic disorders. Computed tomography (CT) is widely used to evaluate cranial and sutural deformities; however, its ionizing radiation is contraindicated in children without significant abnormalities. Magnetic resonance imaging (MRI) offers radiation free scans with superior soft tissue contrast, but unlike CT, MRI cannot elucidate cranial sutures, estimate skull bone density, or assess cranial vault growth. This study proposes a deep learning driven pipeline for transforming T1 weighted MRIs of children aged 0.2 to 2 years into synthetic CTs (sCTs), predicting detailed cranial bone segmentation, generating suture probability heatmaps, and deriving direct suture segmentation from the heatmaps. With our in-house pediatric data, sCTs achieved 99% structural similarity and a Frechet inception distance of 1.01 relative to real CTs. Skull segmentation attained an average Dice coefficient of 85% across seven cranial bones, and sutures achieved 80% Dice. Equivalence of skull and suture segmentation between sCTs and real CTs was confirmed using two one sided tests (TOST p < 0.05). To our knowledge, this is the first pediatric cranial CT synthesis framework to enable suture segmentation on sCTs derived from MRI, despite MRI's limited depiction of bone and sutures. By combining robust, domain specific variational autoencoders, our method generates perceptually indistinguishable cranial sCTs from routine pediatric MRIs, bridging critical gaps in non invasive cranial evaluation.

Figures (3)

• Figure 1: Proposed framework: A VAE generates synthetic CT from input MRI, optimized with adversarial and 3D reconstruction losses. The synthesized CT is processed by a VAE-based, atlas-informed segmentation module that outputs probabilistic segmentation labels.
• Figure 2: Top: Axial, sagittal, and coronal slices for MRI, ground truth CT, and synthesized CT (sCT) using different methods (Ours, nnUNet, 3DCGAN). Bottom: 3D ground truth and predicted cranial bone and suture segmentations for each method. Age = 683 days, Sex = Female
• Figure 3: Axial, sagittal, and coronal views show suture probability heatmaps along with 3D visualization of GT suture segmentation (brown) and predicted segmentation (cyan) overlaid for two samples. (A) Age = 683 days, Sex = Female (B) Age = 80 days, Sex = Female