LUNA16 Lesion Segmentation with U-Net and CeNN Preprocessing

Lung nodule segmentation on CT scans is a key step in computer-assisted analysis for early lung cancer screening. Although U-Net and related deep learning models perform well on many medical imaging tasks, they still struggle with nodules that are exceedingly small or have weak contrast compared to surrounding tissue. In this study, we compare two processing pipelines on the LUNA16 dataset. The first trains a standard 2D U-Net on windowed CT slices. The second adds a Cellular Neural Network (CeNN) that produces two contrast-enhanced versions of each slice. These are stacked with the original image and fed as a three-channel input to the same U-Net architecture. Both pipelines use the same patient-wise data split and training settings so that any differences can be attributed to the processing step. Performance is measured with Dice, IoU, sensitivity, and specificity, and a paired Wilcoxon signed-rank test is applied to the per-patient Dice scores. In the experiments, the CeNN + U-Net variant does not outperform the baseline: Dice and IoU are slightly lower, and the p-value (0.176) indicates that the difference is not statistically significant. These results suggest that, in the simple fixed-template form used here, CeNN preprocessing does not reliably improve lung nodule segmentation, but it provides a useful case of how biologically inspired filters should be evaluated before being added to deep learning pipelines.