(Sunday, 7/14/2019) 4:00 PM - 4:30 PM
Room: Exhibit Hall | Forum 8
Purpose: Convolution Neural Networks (CNN) can be used to accurately segment the liver in MRI images for fully automated liver iron quantification analysis via T2* relaxometry.
Methods: 2D SPGR multi echo (16 echoes) MRI for T2* relaxometry were acquired on 36 patients with hemochromatosis for iron quantification analysis. Liver in 176 slices was segmented in ITK-SNAP and fed into a CNN for segmentation modeling. The network architecture favored for the segmentation task is a a batch norm 512 feature bottleneck U-Net(2). The U-Net was expanded to 16 channel depth to incorporate all 16 acquisition echoes. The model is trained with RMSProp optimizer with Dice similarity coefficient loss over 40 epochs in pytorch. The liver segmentations are applied to collected relaxometry data to calculate liver T2* value. The calculated T2* value is applied to calibration curves from Wood et al to yield liver iron concentration. Liver iron concentration is tested over 90 test slices between manually segmented and U-Net segmented images with paired t test.
Results: Performance metrics are run on a holdout test set of 90 slices. The calculated dice score between hand segmented and CNN segmented was 0.972 ± 0.019. The agreement in liver iron concentration values between hand drawn and convolution network segmented values on the holdout test set was in good agreement(t=3.1, p=0.0023)
Conclusion: Wide 16 channel segmentation neural networks provide excellent models for segmentation. Fully automated deep learning based segmentations achieve a high level of accuracy when compared with hand drawn segmentations. Segmentation of liver has been implemented for automated calculation of liver iron quantification.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
