Author: Marissa Brown, Geoffrey D. Clarke, Luke Norton 👨🔬
Affiliation: University of Texas Health Science Center at San Antonio 🌍
Purpose: To evaluate how different learning strategies affect convolutional neural network (CNN) estimates of the liver's intravoxel incoherent motion (IVIM) parameters.
Methods: A 3-stage U-Net was trained via (i) supervised learning or (ii) transfer learning. For supervised learning, 500 diffusion-weighted images (SNRb0 = 20) were simulated and split 90/10 into training and testing sets. IVIM parameter maps were normalized before network input, and the loss function minimized the mean square error (MSE) between the predicted and ground truth parameters. For transfer learning, the same architecture and hyperparameters were used, but a physics-informed loss minimized the MSE between the input and predicted diffusion signals. The weights/biases of the bridge layer were frozen, and fine-tuning was performed on in vivo data. Performance was assessed using the relative root mean squared error (rRMSE) on simulated data and differences in IVIM parameters in groups of metabolic dysfunction-associated steatotic liver disease (MASLD) and metabolic associated steatohepatitis (MASH) subjects.
Results: On simulated data, the supervised model’s rRMSE values were for the pseudo-diffusion coefficient, D*=0.16, for the perfusion fraction, f=0.018, and for the pure diffusion coefficient, D=0.037. The transfer learning model’s rRMSE values were D*=0.51, f=0.051, and D=0.22. For in vivo data, the supervised model showed significantly lower D in MASH versus non-MASLD subjects (p=0.026). The transfer learning model revealed significant differences in D between non-MASLD and MASLD (p=0.014) as well as between non-MASLD and MASH (p=0.002).
Conclusion: To our knowledge, this is the first study to apply a transfer learning approach to train a CNN for estimating intravoxel incoherent motion parameters. Although the supervised model achieved lower rRMSE when tested on simulated images, the transfer learning approach more effectively distinguished IVIM parameter differences among MASLD subjects. Transfer learning may be a useful learning strategy when datasets are limited.