Calibration and prediction of results after failed injection in SPECT renal dynamic imaging

To predict bilateral renal curves and glomerular filtration rate (GFR) following unsuccessful tracer injection in SPECT renal dynamic imaging using a Transformer model, thereby obviating the need for repeat examinations. We retrospectively studied patients who underwent repeat imaging post-extravasation (2015-2025). Patient height, weight, serum creatinine, urea, uric acid, and renal dynamic curves were used as inputs to develop a Transformer deep learning model for calibrating curves and predicting GFR. Following injection extravasation during SPECT renal dynamic imaging, the proposed model achieved a high overall alignment with the ground truth curves (Median R²>0.93). However, performance variability was observed, with a small subset of outliers showing poor fit (minimum R²<0) due to complex noise artifacts. For the left kidney, the model achieved an R² of 0.9288, with 91.78% of predictions falling within a clinically acceptable tolerance of ±5 ml/min of the ground truth. Similarly, for the right kidney, the R² reached 0.9216, with an accuracy rate of 93.15% based on the same tolerance threshold. It is important to note that the ±5 ml/min criterion was used strictly as a statistical threshold to define 'clinical accuracy' and did not involve any modification of the predicted values. In conclusion, a Transformer model can accurately calibrate and predict bilateral renal curves and GFR values after injection extravasation in SPECT renal dynamic imaging, potentially enhancing clinical workflow efficiency and diagnostic accuracy.