Accelerating imaging: deep learning for enhanced ¹²³I-ioflupane SPECT efficiency

Conventional ¹²³I-ioflupane dopamine-transporter SPECT requires 25-40 min of acquisition, causing patient discomfort and limiting throughput. This study assessed whether deep-learning (DL) reconstruction can yield diagnostic-quality images from a 5-min scan.

We retrospectively analysed 207 studies (1035 slices) obtained between April 2018 and June 2020. After cropping to 64 × 64 striatal regions, 600, 185 and 250 images from 120, 37 and 50 patients were used for training, validation and testing. Six convolutional architectures-U-Net (one-five depths), V-Net, U-Net +  + , R2U-Net, Attention U-Net and TransUNet-were trained to translate 5-min into virtual 25-min images. Image quality was assessed with peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM), analysed by Friedman and Dunn-Holm tests. A blinded reader study involved three nuclear medicine physicians grading 50 cases (100 striata) on a four-point scale; agreement with the 25-min consensus was measured by weighted κ and intra-/inter-observer intraclass correlation coefficients (ICC).

All DL reconstructions significantly outperformed raw 5-min images in PSNR and SSIM (p < 0.01). The four-layer U-Net achieved the highest quality (PSNR 32.7 ± 1.7 dB, SSIM 0.842 ± 0.069), ≈1.8 dB and 0.13 higher than baseline, and statistically indistinguishable from 25-min images (p > 0.05). Reader concordance improved from fair with baseline (κ = 0.29-0.41) to substantial with the four-layer U-Net (κ = 0.62-0.70); intra-reader ICC was 0.84-0.93 and inter-reader ICC 0.73-0.75.

A compact four-layer U-Net restores diagnostic fidelity to 5-min ¹²³I-ioflupane SPECT, enabling an 80% reduction in scan time without loss of quantitative metrics or interpretability. DL-accelerated protocols may enhance comfort, reduce motion artefacts and increase throughput, warranting prospective multicentre validation.