LatXGen: Radiation-Free Sagittal Spine Imaging

• LatXGen is a cross-modal framework that generates lateral spinal X-ray images from posterior RGBD inputs, enabling radiation-free sagittal evaluation in AIS.
• It employs a dual-stage design with Spine Morphology Estimation and Lateral Radiograph Synthesis, using anatomical landmarks for enhanced structural accuracy.
• Empirical results on a large paired dataset show improved performance over baseline GANs, while highlighting challenges like error propagation and dataset diversity.

LatXGen is a cross-modal, dual-stage generative framework that synthesizes realistic lateral spinal radiographs from posterior Red-Green-Blue and Depth (RGBD) images of unclothed backs, with the specific clinical goal of enabling radiation-free quantitative assessment of sagittal spinal alignment in adolescents with Adolescent Idiopathic Scoliosis (AIS). It addresses two coupled difficulties: inferring sagittal spinal morphology from posteroanterior surface geometry, and translating RGBD input into the radiographic domain. The framework first estimates an intermediate lateral spine curve map and then synthesizes the final radiograph, and it is presented as a response to the gap between coronal-plane progress and the relative lack of reliable non-ionizing sagittal evaluation in AIS (Zhao et al., 29 Sep 2025).

1. Clinical problem and task formulation

AIS is treated as a complex three-dimensional spinal deformity rather than a purely coronal-plane curvature problem. The sagittal profile—thoracic kyphosis, lumbar lordosis, and sacral orientation—matters because coronal and sagittal deformities are coupled, and treatment decisions require a comprehensive understanding of both. Standard assessment relies on radiographs, but repeated imaging in adolescents raises concerns about cumulative radiation exposure. Traditional non-radiographic tools such as the forward-bending test, scoliometer measurements, and Moiré topography have utility, but their limitations are described in terms of accuracy, reliability, and observer dependence. Optical sensing, especially RGBD imaging and LiDAR-like depth capture, is positioned as a practical non-ionizing alternative for back-surface acquisition, but prior computational methods are characterized as relying on manual processing or focusing almost entirely on coronal deformity (Zhao et al., 29 Sep 2025).

Within this setting, LatXGen is formulated as a cross-modal novel-view synthesis problem: generate a lateral spinal X-ray image directly from posterior RGBD images of unclothed backs. The objective is not merely visual translation. The stated goal is accurate, radiation-free estimation of sagittal spinal alignment, so the generated lateral radiograph is intended to preserve morphology relevant for downstream quantitative analysis.

2. Input representation and dual-stage design

The pipeline begins with posterior RGB and depth images,

$$I_{RGB} \in \mathbb{R}^{H\times W\times 3}, \qquad I_D \in \mathbb{R}^{H\times W},$$

and aims to generate a lateral spinal X-ray image

$I_X \in \mathbb{R}^{H\times W}.$

A key auxiliary input is a set of 3D anatomical landmarks $v$, extracted from the RGBD data using the authors’ previously developed model. The pipeline first predicts 2D anatomical landmarks directly from RGBD images and then projects them into 3D using the depth map, yielding 3D back-surface keypoints that serve as geometric priors.

A data transformation step rotates the RGB image, depth image, and 3D landmarks by an angle $\theta$ to form an intermediate viewpoint closer to the target lateral perspective: $$\left\{ I_{RGB}^\theta, I_{D}^\theta, v^\theta \right\} = M_T\left( \left\{ I_{RGB}, I_{D}, v \right\} \right).$$ This transformation is motivated by the difficulty of predicting a lateral view directly from a pure posterior view. In the reported ablation, $\theta=45^\circ$ gives the best spine-curve generation performance among the tested angles.

The architecture is explicitly dual-stage. The first stage, Spine Morphology Estimation (SME), predicts an intermediate lateral spine curve map $\hat I_S$. The second stage, Lateral Radiograph Synthesis (LRS), uses the transformed RGBD inputs together with $\hat I_S$ to generate the final lateral radiograph. This progression from structural estimation to image synthesis encodes the central hypothesis that explicit estimation of lateral spinal morphology before radiograph generation improves anatomical consistency and clinical usefulness (Zhao et al., 29 Sep 2025).

3. Spine Morphology Estimation

SME predicts a lateral spine curve map rather than a radiograph: $$\hat I_S = M_{SME}\left(\{I_{RGB}^{\theta}, I_{D}^{\theta}, v^{\theta}\}; \Theta_{SME}\right).$$ Its ground-truth target $I_S$ is obtained by segmenting the true lateral radiograph with SpineHRNet+. The stage comprises six stacked Attention-based Residual Fast Fourier Convolution blocks (ARFFC), followed by a Spatial Deformation Network (SDN) and a decoder (Zhao et al., 29 Sep 2025).

Feature extraction begins with a two-convolution-layer encoder $I_X \in \mathbb{R}^{H\times W}.$0 applied to the transformed RGBD images, producing

$I_X \in \mathbb{R}^{H\times W}.$1

where $I_X \in \mathbb{R}^{H\times W}.$2 and $I_X \in \mathbb{R}^{H\times W}.$3 are global and local feature maps. In parallel, the transformed 3D landmarks are encoded by a two-layer fully connected network $I_X \in \mathbb{R}^{H\times W}.$4: $I_X \in \mathbb{R}^{H\times W}.$5 These landmark features are reused across all ARFFC blocks as anatomy-aware guidance.

Each ARFFC block contains two concatenated AFFC modules with explicit local and global branches. The local branch is updated as

$I_X \in \mathbb{R}^{H\times W}.$6

where $I_X \in \mathbb{R}^{H\times W}.$7 denotes batch normalization plus ReLU. The global branch is updated as

$I_X \in \mathbb{R}^{H\times W}.$8

with $I_X \in \mathbb{R}^{H\times W}.$9 denoting the spectral transform. In the spectral transform, global features are mapped into the frequency domain, processed into key $v$0 and value $v$1, and modulated by landmark-derived query $v$2 through scaled dot-product attention: $v$3 The attended feature is then transformed back: $v$4 The intended effect is simultaneous modeling of local spatial detail, long-range frequency-domain structure, and landmark-guided anatomical constraints.

After the final ARFFC output $v$5, SME applies the SDN: $v$6 The SDN is inspired by spatial transformer networks and uses deformable convolution together with learned sampling parameters to compensate for the fact that even after rotation the input is not a true lateral view. The stated role of the SDN is to improve local curve accuracy and robustness to spatial misalignment.

4. Lateral Radiograph Synthesis and optimization objectives

LRS generates the final radiograph-like lateral X-ray from the transformed optical inputs and the predicted spine structure: $v$7 Architecturally, LRS is also GAN-based. After an encoder, the generator applies six FFC residual blocks and then a decoder: $v$8 where $v$9 denotes the six stacked FFC residual blocks. The FFC blocks follow the configuration of Suvorov et al. The rationale is that once a plausible lateral spine curve map is available, the remaining translation to the radiographic domain is more constrained anatomically (Zhao et al., 29 Sep 2025).

Both SME and LRS are adversarially trained with PatchGAN discriminators. For SME, the discriminator distinguishes real and predicted spine curve maps, and the generator is trained with adversarial and $\theta$0 terms: $\theta$1

$\theta$2

$\theta$3

$\theta$4

For LRS, the losses comprise adversarial supervision, reconstruction, and a spine landmark supervision loss (SLS): $\theta$5

$\theta$6

$\theta$7

Here $\theta$8 is the feature extractor part of a pretrained SpineHRNet+, specifically the layer before the prediction head. SLS acts as a task-specific anatomical consistency constraint rather than a generic perceptual loss.

Training is performed separately for the two stages rather than jointly. Both use Adam with default settings for 400 epochs and batch size 10. The learning rate starts at $\theta$9 and decays to $$\left\{ I_{RGB}^\theta, I_{D}^\theta, v^\theta \right\} = M_T\left( \left\{ I_{RGB}, I_{D}, v \right\} \right).$$0 using cosine decay. RGB and depth images are cropped and resized to $$\left\{ I_{RGB}^\theta, I_{D}^\theta, v^\theta \right\} = M_T\left( \left\{ I_{RGB}, I_{D}, v \right\} \right).$$1, and augmentation includes horizontal flipping, shifting, and random rotations within $$\left\{ I_{RGB}^\theta, I_{D}^\theta, v^\theta \right\} = M_T\left( \left\{ I_{RGB}, I_{D}, v \right\} \right).$$2. The reported implementation platform is an Intel Xeon Platinum 8373C CPU with an NVIDIA RTX 3090 GPU.

5. Dataset, supervision, and empirical performance

A major contribution is the first large-scale paired dataset for this task, comprising 3,264 paired posterior RGBD back-surface images and standing lateral spinal radiographs collected prospectively at Queen Mary Hospital and Duchess of Kent Children’s Hospital in Hong Kong between October 2019 and August 2023 (Zhao et al., 29 Sep 2025). Participants were adolescents aged 10–18 diagnosed with AIS and were predominantly female (72.97%). Exclusion criteria were psychological disorders, trauma affecting posture or mobility, severe dermatological conditions interfering with optical imaging, and known oncological conditions. The split is 2,611 training pairs and 653 testing pairs.

Senior spinal surgeons manually annotated the C7 spinous process and the tip of coccyx on both RGBD images and radiographs to define spinal range and align modalities. The lateral radiographs were also annotated with thoracic kyphosis angle (TKA), lumbar lordosis angle (LLA), and sacral slope angle (SSA). Their definitions are specified as follows: TKA is measured between the T5 superior endplate and T12 inferior endplate; LLA between the L1 superior endplate and S1 superior endplate; SSA between the horizontal plane and the superior endplate of S1.

The main comparative baselines are Pix2PixGAN and Pix2PixHDGAN. For SME, the evaluation metrics are Accuracy, Precision, Sensitivity, F1-score, and IoU. The reported results are:

• Pix2PixGAN: Accuracy 0.973, Precision 0.653, Sensitivity 0.642, F1 0.646, IoU 0.495
• Pix2PixHDGAN: Accuracy 0.978, Precision 0.690, Sensitivity 0.663, F1 0.674, IoU 0.525
• SME without SDN: Accuracy 0.982, Precision 0.750, Sensitivity 0.710, F1 0.727, IoU 0.586
• SME with SDN: Accuracy 0.982, Precision 0.757, Sensitivity 0.712, F1 0.732, IoU 0.592

For LRS, the evaluation metrics are LPIPS, FID, and PSNR. The best-performing full pipeline achieves LPIPS 0.133, FID 93.970, and PSNR 20.854. The comparison points are:

• LatXGen with Pix2PixGAN replacing LRS: LPIPS 0.197, FID 98.711, PSNR 18.837
• LatXGen with Pix2PixHDGAN replacing LRS: LPIPS 0.188, FID 96.970, PSNR 19.071
• Standalone Pix2PixGAN: LPIPS 0.205, FID 140.895, PSNR 18.870
• Standalone Pix2PixHDGAN: LPIPS 0.272, FID 149.138, PSNR 16.153

Task	Best LatXGen result	Comparison point
SME	Accuracy 0.982, Precision 0.757, F1 0.732, IoU 0.592	Pix2PixHDGAN: Accuracy 0.978, Precision 0.690, F1 0.674, IoU 0.525
LRS	LPIPS 0.133, FID 93.970, PSNR 20.854	Pix2PixHDGAN-replaced LRS: LPIPS 0.188, FID 96.970, PSNR 19.071

Clinical parameter prediction is evaluated by applying SpineHRNet+ to the synthesized lateral radiographs. The reported correlations and mean prediction errors are:

• TKA: $$\left\{ I_{RGB}^\theta, I_{D}^\theta, v^\theta \right\} = M_T\left( \left\{ I_{RGB}, I_{D}, v \right\} \right).$$3, mean prediction error $$\left\{ I_{RGB}^\theta, I_{D}^\theta, v^\theta \right\} = M_T\left( \left\{ I_{RGB}, I_{D}, v \right\} \right).$$4, 95% CI $$\left\{ I_{RGB}^\theta, I_{D}^\theta, v^\theta \right\} = M_T\left( \left\{ I_{RGB}, I_{D}, v \right\} \right).$$5
• LLA: $$\left\{ I_{RGB}^\theta, I_{D}^\theta, v^\theta \right\} = M_T\left( \left\{ I_{RGB}, I_{D}, v \right\} \right).$$6, mean prediction error $$\left\{ I_{RGB}^\theta, I_{D}^\theta, v^\theta \right\} = M_T\left( \left\{ I_{RGB}, I_{D}, v \right\} \right).$$7, 95% CI $$\left\{ I_{RGB}^\theta, I_{D}^\theta, v^\theta \right\} = M_T\left( \left\{ I_{RGB}, I_{D}, v \right\} \right).$$8
• SSA: $$\left\{ I_{RGB}^\theta, I_{D}^\theta, v^\theta \right\} = M_T\left( \left\{ I_{RGB}, I_{D}, v \right\} \right).$$9, mean prediction error $\theta=45^\circ$0, 95% CI $\theta=45^\circ$1

For abnormal sagittal alignment classification, the reported values are:

• TKA: sensitivity 0.83, specificity 0.90
• LLA: sensitivity 0.84, specificity 0.86
• SSA: specificity 0.73

The ablation studies reinforce the main design choices. For the transformation angle, $\theta=45^\circ$2 yields the best SME performance:

• $\theta=45^\circ$3: Accuracy 0.965, Precision 0.790, Sensitivity 0.550, F1 0.649, IoU 0.492
• $\theta=45^\circ$4: Accuracy 0.970, Precision 0.613, Sensitivity 0.698, F1 0.631, IoU 0.497
• $\theta=45^\circ$5: Accuracy 0.982, Precision 0.757, Sensitivity 0.712, F1 0.732, IoU 0.592
• $\theta=45^\circ$6: Accuracy 0.975, Precision 0.609, Sensitivity 0.714, F1 0.655, IoU 0.502

For SLS, the comparison is:

• with SLS: LPIPS 0.133, FID 93.970, PSNR 20.854
• without SLS: LPIPS 0.150, FID 95.768, PSNR 19.315

6. Interpretation, scope, and limitations

LatXGen is presented clinically as a promising tool for radiation-free sagittal spine assessment in AIS, particularly in settings that require repeated follow-up. Its framing is broader than screening alone: by generating radiograph-comparable lateral images from posterior RGBD acquisition and enabling estimation of TKA, LLA, and SSA, it is intended to support comprehensive AIS evaluation and longitudinal monitoring while complementing existing radiation-free coronal assessment methods (Zhao et al., 29 Sep 2025).

Several limitations are stated explicitly. The first is error propagation: if SME underestimates or overestimates the spine curve, those errors can propagate into the final synthesized radiograph and reduce anatomical fidelity. The second is metric specificity: standard image synthesis metrics can be influenced by non-spinal regions, which is why they are complemented by clinical sagittal parameter analysis. The third is anatomical difficulty in the sacral region, reflected in the reduced specificity for SSA, which is described as likely due to greater soft-tissue interference around the sacrum.

The generalizability constraints are also clear. The dataset is predominantly female and collected from only two sites within the same geographic and institutional context. Future validation is identified as necessary on broader populations, more diverse clinical settings, and varied imaging protocols, and extension to other spinal pathologies such as Scheuermann’s disease is proposed. The input modality is an unclothed back image, which implies a practical acquisition constraint; the paper does not experimentally test clothed scenarios. It also does not provide an extensive uncertainty analysis, probabilistic confidence estimation, or a systematic catalog of failure cases.

A common overstatement would be to interpret LatXGen as a replacement for conventional radiography in all clinical contexts. That conclusion is not claimed. The method is presented as a promising, radiation-free solution for sagittal spine assessment and as an advance toward comprehensive AIS evaluation, but not as a system already established for preoperative planning or for replacing formal diagnostic radiographs in all scenarios. A plausible implication is that its main immediate significance lies in longitudinal monitoring and repeated assessment, where the reduction of ionizing radiation exposure is especially consequential.