Automated Measurement of Pancreatic Fat and Iron Concentration Using Multi-Echo and T1-Weighted MRI Data

We present an automated method for estimation of proton density fat fraction and iron concentration in the pancreas using both structural and quantitative imaging data present in the UK Biobank abdominal MRI acquisition protocol. Our method relies on automatic segmentation of 3D T1-weighted MRI data using a convolutional neural network and extracting the location of the multi -echo slice through the segmented volume. We finally estimate the fat and iron content in the pancreas using the extracted segmentation as a mask on the multi-echo data. Our segmentation model achieves a mean dice similarity coefficient of 0.842±0.071 on unseen data, which is comparable to the current state of the art for 3D segmentation of the pancreas. The proposed method is efficient and robust and enables an enhanced analysis of spatial distribution of proton density fat fraction and iron concentration over the current practice of manually placing regions of interest on often ambiguous multi-echo data.