U-Net Based Denoising Autoencoder to Half Acquisition Time of Molecular Breast Imaging
Abstract
Purpose
Molecular breast imaging (MBI) with Tc-99m sestamibi and a dedicated gamma camera is useful for dense breast screening, however, a bilateral 2-view (CC and MLO) exam at 10 min/view requires 40 minutes. We aim to halve MBI’s acquisition time through a U-Net based denoising autoencoder.
Methods
Scans from 6086 MBI patients from 2019-2023 were assigned by random stratification to model training (80%), fine-tuning (10%), and final testing (10%). Each MBI view was acquired as ten, 1-minute frames, which, summed together, produced the final image. To train the model, we input summed sets of five frames from each view (frames 1-5; 2-6; 3-7; and 1-4+6). Model performance was assessed with structural similarity index measurements (SSIM) and peak signal-to-noise ratio (PSNR). In 49 patients with diverse clinical presentations, two breast radiologists compared denoised “5 min/view” (frames 1-5) and original 10 min/view versions, while blinded to truth. On a Likert-scale, they answered: "Which study would you prefer to read?", and "Which study do you think is real (10 min/view study)?". Answers were evaluated via Wilcoxon signed rank test and inter-rater variability was measured via Cohen’s Kappa.
Results
Average SSIM and PSNR were 0.917 (SSIM 1.0 = identical images) and 32.6 dB, respectively, in both the fine-tuning and final test set. Readers had a modest preference for the 10 min/view study (p=0.03), with moderate inter-rater agreement (kappa=0.43). Of 49 cases, the two readers had no preference in 29 (59%) and 21(43%). When asked to identify the real 10 min/view images, readers’ confidence was skewed incorrectly toward predicting that the 5 min/view denoised exam was the original exam, however inter-rater agreement was low (kappa=0.004).
Conclusion
The present denoising autoencoder shows promise for halving MBI acquisition time. Future studies will evaluate the impact of denoised exams on lesion detection and clinical management.