Info
Background:
Saturation transfer (ST) MRI is a molecular imaging technique capable of detecting low-concentration metabolites and compounds that are typically invisible to conventional MRI1. Previous ST studies have shown promise in characterizing neurological disorders but interpreting raw ST contrast requires substantial expertise. The application of deep learning to automatic disease detection in ST MRI has remained limited to a small number of studies, often focused on a single neurological disorder, primarily due to the lack of large, annotated ST datasets2,3. Unsupervised anomaly detection (UAD) is an increasingly explored concept in medical image analysis4. UAD eliminates the need for large, heterogeneous data curation by leveraging deep learning to learn distributional properties associated with a single (healthy) class5. By comparing unseen data with the normative model representation from the same subject, anomalies can be discovered. Although applying such methods in conventional MRI has demonstrated promise, these algorithms often rely on T1w hyper-intensity and struggle in other cases6. ST-MRI may be well suited for UAD because deviations from expected molecular exchange effects are directly encoded in the signal yet remain invisible in anatomical MRI.
Hypothesis/Objective:
This work aimed to develop a UAD deep learning framework that combines rapid ST protocols with a physics-aware vision transformer7, and diffusion noise propagation for multi-pathology automatic detection. We postulated that combining ST molecular information encoding with efficient parameter embedding and a vision transformer would enable the detection of challenging neuropathologies.
Methods:
A total of 82 mice, spanning brain cancer, a cuprizone multiple sclerosis (MS) model, MPTP Parkinson’s Disease (PD), Hydrocephalus, and wild-type healthy mice, were scanned using four rapid ST acquisition protocols (targeting MT, amide, rNOE, and guanidinium pools at their respective chemical shifts)8, shortened to 21 s each, at a Bruker 7T scanner. The training used only wild-type (healthy) mice, while all other pathological cases and nineteen healthy controls were split between the validation and test sets. An AI pipeline was designed to learn the robust healthy mouse tissue response to radio-frequency excitation across different ST acquisition protocols (Fig 1). Forward diffusion noise was applied to one of four non overlapping quarters of each image in the input ST-MRI sequence during training to encourage robust reconstruction and simulate realistic signal variability9. A physics aware vision transformer received m=6 raw ST images together with diffusion timestep information and explicit acquisition parameter embeddings (including B₁ amplitude, frequency offset, saturation duration, flip angle, and recovery time).7,10. The transformer was trained to denoise the image series and reconstruct the healthy brain. During inference (t = tmax/4), each quarter was sequentially noised, denoised, and recombined by simple spatial recomposition to generate a full normative ST image series(Fig 2). Normative images were created by subtracting the reconstructed images from the raw input images using the L1 metric (Fig. 3), and an anomaly score per each scan and protocol was obtained by averaging the NRMSE values across the respective m image series.
Results:
The anomaly scores successfully classified healthy vs. pathological mice (Fig. 4), with AUROCs of 0.97 and 0.95 for the semisolid MT and amide-encoding acquisition protocols, respectively (Table 1), which showed the highest sensitivity among the four rapid ST acquisitions. The anomaly maps provided spatially meaningful patterns and insights (Fig. 5), demonstrating localization of anomalies in the corpus callosum and striatum for MS and PD, respectively, in agreement with previous literature11,12. Importantly, these localized effects were absent in conventional anatomical MRI (Fig. 5, center).
Conclusion:
The results demonstrate that efficient molecular data encoding, when fused with diffusion and a transformer-based framework, can yield a robust anomaly score in ST-MRI, offering an annotation-free alternative to supervised and pixel-level approaches13, with only a fraction of traditional acquisition time.
Author
Sahar Ifrah
Tel Aviv University