上海科技大学知识管理系统

In arterial spin labeling (ASL), the flow-sensitive alternating inversion recovery (FAIR) sequence with multi-TI acquisition enables accurate renal perfusion measurement. However, perfusion quantification from ASL signals remains challenging due to the low signal-to-noise ratio (SNR) and its inherent inverse problem. Traditional method for quantification relies on two model-fitting steps. In this study, we propose a BiLSTM-based deep learning (DL) approach trained on simulated pixel-wise ASL signals for simultaneous perfusion quantification and macroscopic vessel exclusion. The network features two decoders for two tasks, with one estimating perfusion along with bolus arrival time (BAT) and bolus length (BL). Compared to the traditional method, the DL method demonstrated superior noise robustness. In simulation experiments, for cortex data with noise standard deviation (STD) of 2% times the mean control signal intensity (or noise level of 2%), the DL method had normalized error of -0.10±0.18, significantly lower than the traditional method’s 0.56±0.56 (P<0.001). For medulla data, the normalized error was 1.07±1.68 (DL) versus 5.19±3.20 (traditional, P<0.001). In in vivo experiments on 16 kidneys from 8 human subjects, similar trends were observed, with DL showing lower normalized error for both renal cortex (0.20±0.48 vs. 0.27±0.50, P<0.001) and medulla (0.29±0.64 vs. 0.58±1.29, P<0.001) at noise level of 2%. The DL-excluded macroscopic vessels closely matched manual exclusion, with intersection over union (IoU) of 80%±6% across 16 kidneys. Perfusion maps estimated by both DL and traditional methods, showed a significant decrease in mean perfusion after DL exclusion (P<0.001). Additionally, BAT and BL quantification were comparable between both DL and traditional methods. In conclusion, the proposed DL method quantified renal perfusion with excellent noise robustness while effectively excluding macroscopic vessels to remove abnormal perfusion values. Along with the perfusion quantification, the DL method also demonstrated its feasibility in BAT and BL quantification.