LAUSR

Abstract Purpose This study aimed to evaluate the accuracy of deep learning (DL)‐based computed tomography (CT) ventilation imaging (CTVI). Methods A total of 71 cases that underwent single‐photon emission CT 81mKr‐gas ventilation (SPECT V) and CT imaging were included. Sixty cases were assigned to the training and validation sets, and the remaining 11 cases were assigned to the test set. To directly transform three‐dimensional (3D) CT (free‐breathing CT) images to SPECT V images, a DL‐based model was implemented based on the U‐Net architecture. The input and output data were 3DCT‐ and SPECT V‐masked, respectively, except for whole‐lung volumes. These data were rearranged in voxel size, registered rigidly, cropped, and normalized in preprocessing. In addition to a standard estimation method (i.e., without dropout during the estimation process), a Monte Carlo dropout (MCD) method (i.e., with dropout during the estimation process) was used to calculate prediction uncertainty. To evaluate the two models’ (CTVIMCD U‐Net, CTVIU‐Net) performance, we used fivefold cross‐validation for the training and validation sets. To test the final model performances for both approaches, we applied the test set to each trained model and averaged the test prediction results from the five trained models to acquire the mean test result (bagging) for each approach. For the MCD method, the models were predicted repeatedly (sample size = 200), and the average and standard deviation (SD) maps were calculated in each voxel from the predicted results: The average maps were defined as test prediction results in each fold. As an evaluation index, the voxel‐wise Spearman rank correlation coefficient (Spearman r s) and Dice similarity coefficient (DSC) were calculated. The DSC was calculated for three functional regions (high, moderate, and low) separated by an almost equal volume. The coefficient of variation was defined as prediction uncertainty, and these average values were calculated within three functional regions. The Wilcoxon signed‐rank test was used to test for a significant difference between the two DL‐based approaches. Results The average indexes with one SD (1SD) between CTVIMCD U‐Net and SPECT V were 0.76 ± 0.06, 0.69 ± 0.07, 0.51 ± 0.06, and 0.75 ± 0.04 for Spearman r s, DSChigh, DSCmoderate, and DSClow, respectively. The average indexes with 1SD between CTVIU‐Net and SPECT V were 0.72 ± 0.05, 0.66 ± 0.04, 0.48 ± 0.04, and 0.74 ± 0.06 for Spearman r s, DSChigh, DSCmoderate, and DSClow, respectively. These indexes between CTVIMCD U‐Net and CTVIU‐Net showed no significance difference (Spearman r s, p = 0.175; DSChigh, p = 0.123; DSCmoderate, p = 0.278; DSClow, p = 0.520). The average coefficient of variations with 1SD were 0.27 ± 0.00, 0.27 ± 0.01, and 0.36 ± 0.03 for the high‐, moderate‐, and low‐functional regions, respectively, and the low‐functional region showed a tendency to exhibit larger uncertainties than the others. Conclusion We evaluated DL‐based framework for estimating lung‐functional ventilation images only from CT images. The results indicated that the DL‐based approach could potentially be used for lung‐ventilation estimation.