Convolutional neural network for simultaneous assessment of iron overload and fat fraction in pancreatic MRI

Thalassemic patients are characterised by a reduced or impaired synthesis of haemoglobin. The therapy currently in use consists of blood transfusions. In the long term, the latter leads to a clinically relevant accumulation of iron in hepatocytes, in parenchymal cells of the pancreas and myocardium. This work focuses on pancreatic iron overload since it is correlated with the accumulation of iron in the heart and can lead to diabetes mellitus. For the assessment of iron overload, the T2* value is calculated using a fitting method on pancreatic gradient-echo multi-echo MRI. In fact, the T2* value of a tissue is decreased by the presence of iron. The signal is modelled as an exponential with an R2* decay rate (R2*= 1/T2*). Moreover, the fitting method calculates the fat fraction (FF) in order to rectify the estimated T2* value and then make a fair evaluation of pancreatic iron overload. In this work, we propose a convolutional neural network to assess T2* and FF on pancreatic MRI. The network developed is a 10-input 2-output channel U-Net 2D, which builds R2* and FF maps for each pancreatic region. At the end of the test phase, this network was compared with two reference fitting methods: the software HIPPO-MIOT, which is currently in use at “Fondazione G. Monasterio”, and Hernando’s algorithm. The results are very promising since the U-Net gives similar results to the reference fitting algorithms but with the advantage of creating an automatic and faster method