Renal Stone Quantification in Contrast Enhanced CT Using Convolutional Neural Network Assisted Dual-Energy Virtual Non-Contrast Imaging
H Gong*, A Ferrero, J Marsh, N Huber, J Fletcher, C McCollough, S Leng, Mayo Clinic, Rochester, MN
Presentations
TU-D-TRACK 3-4 (Tuesday, 7/27/2021) 2:00 PM - 3:00 PM [Eastern Time (GMT-4)]
Purpose: Virtual non-contrast (VNC) technique has been proposed to detect renal stones in contrast-enhanced dual-energy CT (DECT) but has demonstrated limitations of reduced stone volume and removal of small stones. We aim to improve stone detection and quantification accuracy by developing a novel convolutional neural network (CNN) assisted VNC technique.
Methods: A dual-task CNN with customized loss functions was developed to synergistically integrate material-classification and -decomposition. The classification provided a probability weighting scheme to material decomposition, to minimize the ambiguity between calcium and iodinated soft-tissue. Multi-energy CT phantoms were scanned on a dual source DECT (80/Sn150 kV) and image patches (50x50 pixels) of insert materials (e.g. hydroxyapatite, iodine, and iodinated-blood mixture) were used to train the CNN. Random-shaped numerical inserts with additional densities were simulated to augment training dataset. Iodine component was subtracted from original CT images to generate VNC images. Human stone specimens (n=35; 20 calcium-based/15 non-calcium-based) embedded in an abdomen-sized water phantom were scanned with and without iodine contrast (20mg/cc). CT number and volume per stone was measured using in-house stone quantification software and compared to commercial VNC (Syngo.VIA, Siemens). Mean-absolute-percent-error (MAPE) was calculated using true non-contrast (TNC) images as ground truth and Wilcoxon signed rank test (5% significance level) was used to gauge stone volume discrepancy between VNC and TNC images.
Results: CT number of stones from commercial VNC were significantly reduced (MAPE: 32.9%), resulting in significantly-underestimated stone volume (bias: -26.4%, MAPE 26.5%, p<0.01). Stones were much brighter in CNN-assisted VNC than those in commercial VNC, and density plot showed CT numbers recovered to those of the original CT images. The stone volume was also more accurate in CNN-assisted VNC, with not significant difference from that of TNC (bias: -2.1%, MAPE 10.6%, p=0.568).
Conclusion: The CNN-assisted VNC improves the quantification of renal stones in contrast-enhanced DECT.
Funding Support, Disclosures, and Conflict of Interest: This work was supported by the National Institutes of Health under award numbers of R01 EB016966 and EB028590. NVIDIA Corporation donated Titan V GPU used for this research. Research support is provided to Mayo Clinic from Siemens Healthcare GmbH, unrelated to this work.
Handouts
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email