Purpose: To develop a task-based loss function in a deep convolutional neural network (CNN) to reduce noise and maintain spatial details simultaneously.
Methods: A U-Net CNN was trained with simulated-quarter-dose head CT images reconstructed with smooth and sharp kernels and using routine-dose images of each kernel as targets. A baseline model was trained using mean-squared-error (MSE) loss and image patches from 100 scans; 5 additional scans were used for validation and 5 more for testing. To better retain sharpness in the bone and reduce noise in the brain, a loss function with two task-based loss terms was introduced. The first term enforced similarity with the sharp kernel target images in the HU range outside of the brain (<0 or >80) while penalizing sharp kernel noise within the brain HU range using total variation loss. The second term enforced similarity with the smooth kernel target images within the brain HU range. Results from the task-based model were compared against those of the baseline model using region of interest (ROI) measurements of standard deviation to assess noise reduction and line profile measurements to assess spatial resolution preservation.
Results: The CNN model trained using the task-based loss function achieved a 51.6% and 70.5% greater noise reduction when applied to routine and low-dose CT images compared to the MSE loss model while simultaneously maintaining image sharpness in the bone comparable to the sharp kernel input.
Conclusion: A task-based loss function that specified where each input kernel should be emphasized, outperformed the same model trained under equal conditions but using MSE loss alone. Superior performance was noted in terms of noise reduction and sharpness preservation.