Discriminative Image Enhancement for Robust Cascaded Segmentation of CT Images

Objective: Cascaded/attention-based neural network has become common in image segmentation. This work proposes to improve its robustness by adding discriminative image enhancement to its attention mechanism. Unlike prior work, this image enhancement can also be applied as data augmentation and easily adapted for existing models. Its generalization can improve accuracy across multiple segmentation tasks and datasets. Methods: The method first localizes a target organ in a 2D fashion to obtain a tight neighborhood of the organ in each slice. Next, the method computes an HU histogram of a region combined from multiple 2D neighborhoods. This allows the method to adaptively handle HU-range difference among images. Then, HUs are nonlinearly stretched through a parameterized mapping function providing discriminative features for neural network. Varying the function parameters creates different intensity distribution of the target region. This effectively enhances and augments image data at the same time. The HU-reassigned region is then fed to a segmentation model for training. Results: Our experiments on liver and kidney segmentation showed that even a simple cascaded 2D U-Net model could deliver competitive performance in a variety of datasets. In addition, cross-validation and ablation analysis indicated robustness of the method even when the number of original training samples was limited. Conclusion: With the proposed technique, a simple model with limited training data can deliver competitive performance. Significance: The method significantly improves robustness of a trained model and is ready for generalization to other segmentation tasks and attention-based models. Accurate models can be simpler to save computing resources.