- The paper introduces a text-controllable framework for denoising Positron Emission Tomography (PET) images trained using text embeddings, outperforming baselines with average SSIM and PSNR improvements of 0.075 and 3.7 dB, respectively, at 1/100 full dose.
- The integration of CLIP embeddings of count-level text descriptions into the U-Net architecture enables variable input-output pairings, achieving consistent improvements across and within various dose ranges.
- A notable limitation involves the lack of evaluation outside 2D axial slices and scanning protocols, as well as the necessity of paired low-high-count data for post-processing.
Overview
This paper presents a text-controllable denoising framework for Positron Emission Tomography (PET) that conditions image enhancement on natural-language descriptions of count levels. Rather than training a fixed input–output pair model, the authors integrate CLIP text embeddings of count-level prompts (e.g., "a 1/100 count level PET image") into both the encoder and decoder pathways of a U-Net, allowing a single model to denoise PET images across a wide range of count levels and to target arbitrary higher-count outputs. The work is motivated by clinical scenarios—pediatric imaging, oncology follow-up, dose reduction protocols—in which low-count acquisitions produce noisy images that degrade diagnostic accuracy.
Method
The architecture is a U-Net with a dual text-embedding pathway built on the pretrained CLIP ViT-B/32 text encoder (Radford et al., 2021). In the encoder path, an embedding of the input count description is broadcast and element-wise multiplied with feature maps at every U-Net layer, conditioning feature extraction on the input count context. In the decoder path, a parallel embedding derived from the output (target) count description is injected at each decoder layer to steer reconstruction toward the desired count-level characteristics. The output count level is constrained to be no lower than the input.
Training uses axial slices of low-count PET images paired with their count-level text embedding and a higher-count ground-truth slice. Count conditions are randomly sampled at each training step to encourage generalization across input–output combinations. Optimization uses AdamW (learning rate 0.001, batch size 32) for 500 epochs on two NVIDIA A100 GPUs, with an MSE loss against the full-count reference. The dataset was split 80/20 train/test with patient-level separation to prevent leakage.
Data and evaluation
Experiments use the publicly available Siemens Biograph Vision Quadra total-body dataset from the Ultra-low Count PET Imaging Challenge: 387 18F-FDG scans acquired in list mode, enabling retrospective rebinding into shortened time windows to simulate 1/100, 1/20, 1/10, 1/4, and 1/2 count reconstructions alongside the full-count reference (Yu et al., 28 Feb 2025).
Evaluation proceeds in two parts. First, images at all simulated count levels are denoised to synthetic full-count images; SSIM and PSNR are computed against the real full-dose reference for both the original low-count inputs and the denoised outputs. Across all count levels, the denoised images are consistently closer to the full-count reference than the raw low-count inputs on both metrics, indicating that a single text-conditioned model handles the entire range from 1/100 to 1/2 count without per-level retraining. Second, on the hardest setting (1/100 → full count), the proposed method is compared against U-Net and CycleGAN baselines trained specifically on that fixed pair. The proposed method yields higher PSNR and SSIM than both baselines, with visual results closer to the full-count reference. The authors note this comparison is deliberately limited to basic baselines, since no existing method offers equivalent flexible functionality; the comparison therefore establishes competitiveness rather than state-of-the-art superiority in absolute terms.
The central claim—that semantic conditioning via CLIP embeddings enables one model to span many count levels while outperforming fixed-pair models on their own task—is supported by these results, though the quantitative gains over U-Net/CycleGAN are reported graphically rather than with statistical significance testing.
Limitations and open questions
The paper concedes several constraints. Training depends on paired low-/high-count data generated by retrospective list-mode rebinding; the authors note that access to sufficient raw list-mode data would be needed to simulate truly arbitrary dose levels, and the current study is limited by paired-data availability. Only 2D axial slices are processed, leaving 3D volumetric consistency unaddressed. The evaluation relies on SSIM/PSNR against a single scanner/tracer (18F-FDG, Biograph Vision Quadra), so cross-scanner and cross-tracer generalization remain untested. No lesion-level or diagnostic-task validation is performed, so whether text-guided synthesis preserves quantitative tracer uptake (e.g., SUV) relevant to clinical reading is an open question. Finally, the constraint that output count must exceed input count is imposed rather than learned, and the behavior of the model under out-of-distribution prompt phrasings is not characterized.
Conclusion
This work demonstrates that CLIP-derived count-level semantics can serve as an effective conditioning signal for PET denoising, replacing fixed-pair training with a single flexible model covering count levels from 1/100 to full dose. Quantitative results show consistent SSIM/PSNR improvements over raw low-count inputs across all levels and superior performance relative to U-Net and CycleGAN on the 1/100-count task. The approach points toward count-aware, protocol-adaptive PET reconstruction, contingent on future validation of quantitative accuracy, 3D extension, and generalization beyond the evaluated scanner and tracer.