Patlak-Guided Self-Supervised Learning for Dynamic PET Denoising

Le résumé fourni par la source

Dynamic PET imaging delivers crucial FDG metabolism data for diagnosis and treatment response assessment. Count limitations pose challenges to existing methods for estimating Ki images. We proposed a deep-learning framework that integrates the Patlak model into UNet-LSTM without imposing image size constraints. The study used a dynamic PET dataset with 28 patients, employing a three-fold training scheme. Training from scratch was challenging due to noise and weak supervision, soa pre-trained 3D UNet model on prompts-matched static images from 100 subjects was utilized. The pre-trained model was then used to initialize the downsampling and upsampling blocks of the UNet-LSTM network, facilitating convergence. After weight initialization, two approaches were used: Model A utilized raw dynamic frames as the input, while Model B utilized denoised dynamic frames by the pre-trained model as the input. Evaluation of 28 patients with 54 lesions reveals that the proposed approach achieves superior noise reduction and lesion detectability compared to both the prompts-matched model and direct reconstruction methods. Overall, Model B exhibits the most promising outcomes, as indicated by the enhancement in lesion Signal-to-Noise Ratio from 20.6 ± 13.6 (Indirect Recon), 33.7 ± 25.9 (Prompts-Matched), 34.5 ± 22.2 (Model A) to 47.6 ± 40.2 for 14 -frame prediction and from 20.9 ± 13.5 (Direct Recon), 14.2 ± 9.3 (Indirect Recon), 24.8 ± 18.1 (Prompts-Matched), and 26.9 ± 17.9 (Model A) to 32.9 ± 26.6 for 6 -frame prediction. In conclusion, we propose the Patlak-Guided Self-Supervised learning method, which integrates spatial and temporal data for dynamic PET denoising, enhances noise reduction and preserves lesions. The superior lesion SNR highlights its potential clinical utility and imaging quality improvement.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.