Applying a Deep Learning Model for Total Kidney Volume Measurement in Autosomal Dominant Polycystic Kidney Disease

Background: Autosomal dominant polycystic kidney disease (ADPKD) is the most common hereditary renal disease leading to end-stage renal disease. Total kidney volume (TKV) measurement has been considered as a surrogate in the evaluation of disease severity and prognostic predictor of ADPKD. However, the traditional manual measurement of TKV by medical professionals is labor-intensive, time-consuming, and human error prone.

Materials and methods: In this investigation, we conducted TKV measurements utilizing magnetic resonance imaging (MRI) data. The dataset consisted of 30 patients with ADPKD and 10 healthy individuals. To calculate TKV, we trained models using both coronal- and axial-section MRI images. The process involved extracting images in Digital Imaging and Communications in Medicine (DICOM) format, followed by augmentation and labeling. We employed a U-net model for image segmentation, generating mask images of the target areas. Subsequent post-processing steps and TKV estimation were performed based on the outputs obtained from these mask images.

Results: The average TKV, as assessed by medical professionals from the testing dataset, was 1501.84 ± 965.85 mL with axial-section images and 1740.31 ± 1172.21 mL with coronal-section images, respectively (p = 0.73). Utilizing the deep learning model, the mean TKV derived from axial- and coronal-section images was 1536.33 ± 958.68 mL and 1636.25 ± 964.67 mL, respectively (p = 0.85). The discrepancy in mean TKV between medical professionals and the deep learning model was 44.23 ± 58.69 mL with axial-section images (p = 0.8) and 329.12 ± 352.56 mL with coronal-section images (p = 0.9), respectively. The average variability in TKV measurement was 21.6% with the coronal-section model and 3.95% with the axial-section model. The axial-section model demonstrated a mean Dice Similarity Coefficient (DSC) of 0.89 ± 0.27 and an average patient-wise Jaccard coefficient of 0.86 ± 0.27, while the mean DSC and Jaccard coefficient of the coronal-section model were 0.82 ± 0.29 and 0.77 ± 0.31, respectively.

Conclusion: The integration of deep learning into image processing and interpretation is becoming increasingly prevalent in clinical practice. In our pilot study, we conducted a comparative analysis of the performance of a deep learning model alongside corresponding axial- and coronal-section models, a comparison that has been less explored in prior research. Our findings suggest that our deep learning model for TKV measurement performs comparably to medical professionals. However, we observed that varying image orientations could introduce measurement bias. Specifically, our AI model exhibited superior performance with axial-section images compared to coronal-section images.