Accelerating MRI with Longitudinally-informed Latent Posterior Sampling

TL;DR

This work addresses the challenge of accelerating MRI reconstruction when longitudinal paired training data are scarce. It introduces LAPS, a diffusion-prior framework that learns from unpaired images in latent space and leverages a subject's prior DICOM as an inference-time cue, eliminating the need for longitudinal training pairs. A key contribution is AutoInit, which estimates a projection timepoint $t_p$ and phase to hot-start latent posterior sampling, balancing prior influence with new measurements. Evaluated on the SLAM dataset, LAPS achieves higher acceleration capabilities (up to $1$D: ~9x, $2$D: ~30x) while maintaining image quality and robustness to anatomical changes and misregistration, outperforming both longitudinal and non-longitudinal baselines. The open SLAM dataset and the proposed methodology offer a practical, scalable path toward clinically accelerated, longitudinally-aware MRI reconstruction.

Abstract

Purpose: To accelerate MRI acquisition by incorporating the previous scans of a subject during reconstruction. Although longitudinal imaging constitutes much of clinical MRI, leveraging previous scans is challenging due to the complex relationship between scan sessions, potentially involving substantial anatomical or pathological changes, and the lack of open-access datasets with both longitudinal pairs and raw k-space needed for training deep learning-based reconstruction models. Methods: We propose a diffusion-model-based reconstruction framework that eliminates the need for longitudinally paired training data. During training, we treat all scan timepoints as samples from the same distribution, therefore requiring only standalone images. At inference, our framework integrates a subject's prior scan in magnitude DICOM format, which is readily available in clinical workflows, to guide reconstruction of the follow-up. To support future development, we introduce an open-access clinical dataset containing multi-session pairs including prior DICOMs and follow-up k-space. Results: Our method consistently outperforms both longitudinal and non-longitudinal baseline reconstruction methods across various accelerated Cartesian acquisition strategies. In imaging regions highly similar to the prior scan, we observe up to 10\% higher SSIM and 2 dB higher PSNR, without degradation in dissimilar areas. Compared to longitudinal reconstruction baselines, our method demonstrates robustness to varying degrees of anatomical change and misregistration. Conclusion: We demonstrate that prior scans can be effectively integrated with state-of-the-art diffusion-based reconstruction methods to improve image quality and enable greater scan acceleration, without requiring an extensive longitudinally-paired training dataset.

Figures (17)

• Figure 1: Illustration of our proposed method, LAPS. From left to right: We begin with undersampled k-space from a new scan and retrieve the corresponding prior DICOM scan of the subject. Next, we estimate an initial phase ${\bm{\phi}}$ and find the optimal projection point $t_p$ using $\mathop{\texttt{AutoInit}}\nolimits$. We encode the phase-modulated prior and project it to $t_p$, which provides the initialization for our proposed LDM reconstruction that also incorporates the newly acquired k-space data. Finally, the decoded output of the diffusion process is passed through a few additional DC steps, refining the output image. Compared to LDPS (above), LAPS starts the diffusion process much earlier than $T$ via $\mathop{\texttt{AutoInit}}\nolimits$, runs $n_{\rm{opt}}$ DC steps per iteration rather than 1, and enforces additional DC at the output.
• Figure 2: Comparison of reconstruction methods at varying 1D acceleration rates in a case with minimal longitudinal change. Axial $T_2$-weighted images at the level of the basal ganglia and insula in a patient with an incidentally discovered sub-centimeter cerebellar lesion (not shown), presumed to be a low-grade tumor. The lesion and other brain regions remained stable between the two exams performed six months apart. The first row shows, from left to right, the target (ground truth) image, the prior scan, and the amplified difference between them. The second row displays the undersampling masks used for reconstruction at acceleration rates of $R = 3$, $6$, and $9$. Subsequent rows present reconstructions from different methods, with our proposed method (LAPS) shown in the last row. Each image is split vertically: the left half shows the error map (amplified $\times5$) of the corresponding left half of the reconstruction, while the right half displays the actual reconstructed image. A zoomed-in region is shown to the right of each reconstruction to highlight fine details and local differences. Finally, the PSNR and SSIM of each reconstruction are indicated at the top left corner.
• Figure 3: Comparison of reconstruction methods across varying 2D acceleration rates in a case with visible longitudinal changes. Coronal $T_2$-weighted images from a pediatric brain scan with a low-grade glioma. The initial scan (prior) shows a 6 cm $T_2$-hyperintense mass in the right frontal and parietal lobes. In the follow-up scan (target), the mass has been biopsied, with the biopsy cavity clearly visible. The first row displays (left to right): the follow-up scan (target), the prior scan, and an amplified difference image highlighting changes between them. The second row shows the undersampling masks used in the experiments, corresponding to acceleration rates of $R = 10$, $20$, and $30$. The subsequent rows present reconstructed images from various methods, with LAPS shown in the bottom. The PSNR and SSIM of each reconstruction are indicated at the top right corner. A zoomed-in region around the biopsy cavity is used to assess each method’s ability to accurately reconstruct pathological tissue and capture deviations over time.
• Figure 4: Mean global performance metrics across various 1D and 2D accelerations. For brevity, performance is only shown for LAPS with $\mathop{\texttt{AutoInit}}\nolimits$ and the top 2 performing baselines (CAPS and MODL).
• Figure 5: Patch-based metric analysis. Each slice in the test set was divided into $32 \times 32$ patches with 50% overlap, and the similarity between corresponding patches in the target and prior scans was computed using cosine similarity. To evaluate performance as a function of prior similarity, we aggregated all patches across the test set and grouped them into bins based on similarity percentiles. In each subplot, the $x$-axis represents a similarity percentile bin: the leftmost bin includes the 10% most similar patches, followed by patches in the 10–50% range (still similar but less so), then the 50–90% range, and finally, the rightmost bin includes the 10% most dissimilar patches. The bar labeled Prior shows the metric computed between the target and prior patches within each range.