The effect of preprocessing on convolutional neural networks for medical image segmentation

In recent years, deep learning has become the leading method for medical image segmentation. While the majority of studies focus on developments of network architectures, several studies have shown that non-architectural factors also play a substantial role in performance improvement. An important factor is preprocessing. However, there is no agreement on which preprocessing steps work best for different applications. The aim of this study was to investigate the effect of preprocessing on model performance. To this end, we conducted a systematic evaluation of 24 preprocessing configurations on three clinical application datasets (brain, liver, and knee). Different configurations of normalization, region of interest selection, bias field correction, and resampling methods were applied before training one convolutional neural network. Performance varied up to 64 percentage points between configurations within one dataset. Across the three datasets, different configurations performed best. In conclusion, to improve model performance, preprocessing should be tuned to the specific segmentation application.