Two-Stage Deep Learning Framework Improves Deformable SPECT and CT Registration for Dosimetry

A computational framework combining inter-patient pre-training with intra-patient few-shot adaptation has demonstrated enhanced accuracy and computational efficiency for deformable single-photon emission computed tomography and computed tomography registration in targeted radionuclide therapy. Developed to address severe data limitations in nuclear medicine clinics, the unsupervised deep learning pipeline, termed Nufit-Reg, registered longitudinal multimodal scans in approximately nine seconds while significantly reducing biokinetic curve fitting errors and producing more stable voxel-level tumor absorbed dose distributions. The investigation, validated retrospectively on clinical iodine-131 therapies for differentiated thyroid cancer, demonstrated how structural constraints derived from high-resolution computed tomography anatomical references can guide spatial alignment across serial functional emissions without coupling the network optimization to noisy tracer kinetics. By establishing reliable spatial correspondence across sequential treatment time points, the system provides an automated method for estimating voxel-level time-integrated activity maps, offering an essential technical foundation for individualized treatment verification where large institutional training datasets are completely inaccessible.

Radionuclide therapy has sustained clinical application for over eight decades, beginning with iodine-131 administrations for thyroid disorders and expanding rapidly following recent regulatory approvals of lutetium-177 radiopharmaceuticals. Unlike conventional external beam radiotherapy, systemically administered radiopharmaceuticals achieve cytotoxic cellular damage by selectively depositing high-energy emissions directly into neoplastic tissue, precipitating lethal single-strand and double-strand deoxyribonucleic acid breaks. Establishing accurate dose-response relationships and predicting biological toxicities in healthy organs requires precise estimation of absorbed radiation doses across heterogeneous tumor environments. Voxel-level dosimetry accomplishes this by integrating anatomical boundaries from computed tomography with serial functional emission measurements from quantitative single-photon emission computed tomography across multiple days. However, accurate quantification depends entirely on image registration to align longitudinal acquisitions. Previous experimental evaluations have demonstrated that modest spatial misalignments of less than nine point five millimeters or angular deviations below six degrees can introduce absorbed dose calculation errors reaching ninety percent, which subsequently distort downstream time-activity curve fitting and dose kernel convolution algorithms.

Direct deformable registration based solely on functional emission intensities presents formidable challenges in clinical practice due to progressive tracer decay, dynamic biological clearance kinetics, restricted spatial resolution, and unfavorable signal-to-noise ratios. Conversely, computed tomography provides rigid anatomical references that remain comparatively invariant across longitudinal imaging intervals. Because hybrid systems acquire emission and transmission datasets within a co-registered physical framework, computed tomography spatial consistency can serve as an indirect constraint for tracking underlying functional tracer distributions. Although unsupervised deep learning networks have achieved remarkable computational speed and competitive accuracy in unimodal tasks such as computed tomography to computed tomography or magnetic resonance imaging alignments, their deployment in multimodal longitudinal emission tomography has remained constrained. Most nuclear medicine clinics implement internal protocols encompassing merely ten to twenty patients, creating a profound data scarcity barrier that prevents the clinical translation of automated neural registration tools.

To overcome severe training data constraints, researchers implemented Nufit-Reg using a retrospective single-institution clinical cohort from the First Affiliated Hospital of the University of Science and Technology of China. The patient group comprised adult inpatients aged between eighteen and sixty years with pathologically confirmed differentiated thyroid cancer who had completed total or near-total thyroidectomy and were scheduled for remnant ablation therapy with one hundred millicuries of oral iodine-131, without documented distant metastases. The clinical imaging data were divided into two strictly separated, non-overlapping cohorts at the patient level. The pre-training dataset consisted of fifty-eight patients whose single-photon emission computed tomography and computed tomography volumes were acquired approximately ninety-six hours following therapeutic administration. The longitudinal tuning dataset comprised twelve distinct patients who underwent quantitative serial acquisitions across three sequential post-administration intervals at thirty, fifty-four, and seventy-eight hours, yielding thirty-six intra-patient registration pairs for domain adaptation and cross-validation.

All acquired volumetric data underwent rigorous preprocessing to reconcile disparate native fields of view and voxel resolutions between modalities before network ingestion. Overlapping physical volumes were identified and cropped to ensure matching spatial coordinates while strictly preserving spatial origin and orientation metadata recorded within primary digital imaging headers. Computed tomography intensity values were truncated within a window spanning minus one thousand to two thousand Hounsfield units and scaled to a unit range, whereas emission volumes were normalized through min-max scaling to standardize voxel values across datasets. All image matrices were resampled to a uniform spatial dimension of two hundred fifty-six by two hundred fifty-six by one hundred twenty-eight voxels via linear interpolation, avoiding high-order spatial artifacts. Initial baseline alignment was established through rigid registration of computed tomography volumes using default multi-resolution parameter configurations within the Elastix library, and the resulting affine transformation matrices were simultaneously applied to corresponding emission volumes using linear interpolation to ensure multimodal congruence before deformable modeling.

The neural architecture of Nufit-Reg employs a dual-encoder design paired with a single decoding network that constructs a three-dimensional displacement field. Each encoder accepts a two-channel input representing fixed and moving volumes of a single imaging modality with initial spatial dimensions of two hundred fifty-six by two hundred fifty-six by one hundred twenty-eight voxels and sixty-four baseline feature channels, doubling capacity at each hierarchical depth. Both encoders incorporate Swin Transformer blocks augmented by Fusion Attention Blocks and an Overlapping Attention Block. The fusion units combine convolutional operations with self-attention mechanisms computed within local sub-volumes of seven by six by five voxels, mitigating local redundancy while capturing extended spatial relationships. Overlapping attention blocks enable information exchange across adjacent window boundaries. To facilitate multimodal feature integration, cross-stitch units were embedded at the initial two hierarchical stages and the central bottleneck layer, establishing soft parameter sharing through learnable linear activation combinations. Continuous super-resolution upsampling layers in the decoder reconstruct continuous coordinate displacement fields, mitigating discretization degradation inherent to standard interpolation operations.

The optimization strategy uses a composite unsupervised objective function that isolates structural alignment to the computed tomography domain while enforcing topological plausibility. To avoid unstable training caused by tracer kinetic variations and poor functional signal-to-noise ratios, single-photon emission computed tomography intensity losses were excluded entirely from gradient backpropagation. The overall loss combines computed tomography structural similarity, mutual information, a diffusion regularizer penalizing spatial deformation gradients, and a Jacobian determinant penalty preventing negative local volumes. Training proceeded across two sequential stages using an NVIDIA RTX 3090 graphics processing unit. In the first stage, one hundred fifty inter-patient pairs selected by minimal Euclidean distances between tracheal anatomical boundaries across fifty-eight pre-training subjects were trained to establish general anatomical deformation priors. In the second stage, few-shot fine-tuning transferred the network parameters to twelve intra-patient subjects across five independent patient-level cross-validation folds with online affine data augmentation, reducing the diffusion regularization weight from one point zero to zero point five to permit flexible patient-specific local displacements.

Comprehensive evaluation demonstrated that Nufit-Reg achieved superior image similarity across clinical test cases compared with conventional optimization-based B-spline registration in Elastix and the deep learning baseline UTSRMorph. Evaluated across five validation folds, Nufit-Reg achieved a weighted structural similarity index of zero point eight nine six eight plus or minus zero point zero two four one in the primary abstract report and zero point eight nine two three plus or minus zero point zero three one four in cohort analysis, significantly exceeding Elastix at zero point eight six seven eight and UTSRMorph at zero point eight eight one one. Because the weighted structural similarity metric incorporates voxel weights proportional to localized radiopharmaceutical uptake, these gains indicate precise spatial correspondence within clinically vital high-activity tumor zones. Unweighted structural similarity and local normalized cross-correlation also demonstrated statistically significant enhancements over Elastix with large effect sizes. Computationally, Nufit-Reg completed full volumetric deformation in approximately nine seconds on a single graphics processing unit, contrasting with approximately one minute required by CPU-based iterative execution in Elastix.

Analysis of deformation topology via Jacobian determinants confirmed that Nufit-Reg maintained physical plausibility while accommodating localized physiological distortions. Across all evaluated methods, the fraction of voxels displaying negative Jacobian determinants remained below one in ten thousand, indicating the absence of substantial folding or singular collapses. Nufit-Reg recorded the lowest overall fold ratio at three point eight five times ten to the negative fifth power, while achieving a mean Jacobian determinant of zero point nine nine seven four, closely approaching ideal volume preservation alongside UTSRMorph at zero point nine nine seven zero, compared to zero point nine seven three nine observed for Elastix.

Downstream dosimetric verification, conducted by fitting bi-exponential time-activity curves to tumor voxels and executing GATE Monte Carlo radiation transport simulations, demonstrated tangible improvements in dosimetric reliability. Voxel-level kinetic curve fitting with Nufit-Reg produced a mean coefficient of determination of zero point eight eight five zero and a mean root mean square error of one point seven eight eight times ten to the fifth becquerels per milliliter, representing an eleven point seven percent error reduction compared to Elastix. This reduction in kinetic fitting error directly curbed non-physical dose outliers, shifting maximum D2 percent values from one hundred ten point eight one gray down to ninety-five point one one gray while modestly raising median low-dose D98 percent from three point one eight gray to three point three zero gray. The authors emphasized that because high-resolution in vivo ground truth absorbed doses cannot be directly measured in human subjects, these metrics demonstrate technical consistency and biokinetic stability rather than absolute dosimetric truth, requiring further validation across multi-center cohorts and alternative radiopharmaceuticals.

Subject of Research: Unsupervised deep learning deformable SPECT/CT image registration for voxel-level radionuclide therapy dosimetry

Article Title: Data-efficient unsupervised deep learning deformable SPECT/CT registration framework for voxel-level radionuclide therapy dosimetry: validation using clinical 131I DTC therapy data

Article References: Liu, J., Wang, X., Zhang, H., Xu, Z., Wang, X., & Xu, X. G. (2026). Data-efficient unsupervised deep learning deformable SPECT/CT registration framework for voxel-level radionuclide therapy dosimetry: validation using clinical 131I DTC therapy data. EJNMMI Physics, 13(1), Article 97. https://doi.org/10.1186/s40658-026-00911-2

Image Credits: AI Generated

DOI: 10.1186/s40658-026-00911-2

Keywords: Nufit-Reg, deformable image registration, SPECT/CT, radionuclide therapy, voxel-level dosimetry, iodine-131, differentiated thyroid cancer, unsupervised deep learning, Swin Transformer, cross-stitch units, time-integrated activity, time-activity curve, Monte Carlo simulation, Elastix, UTSRMorph, structural similarity, Jacobian determinant, radiation dose distribution