Cross-Dataset Generalization of RCED-UNet3+ for Robust Lung Nodule Segmentation

Abstract

Aim of the Study: The main objective of this study is to ensure that pulmonary nodules are properly identified on CT scans to detect lung cancer at the earliest stage. In particular, the study aims to assess the generalizability and robustness of the RCED-UNet3+ segmentation model across datasets with varying imaging properties.

Methodology: The current study is a continuation of our earlier work in which the RCED-UNet3 + model has a Dice score of 0.984 on the LIDC-IDRI dataset. The model will be further evaluated on the LUNA16 dataset to extend the results achieved previously. The datasets also vary in terms of scanner technologies and image resolutions, and hence offer a platform to evaluate the model in terms of its consistency in performance in different imaging sources. The p-value analysis (p < 0.05) was used to test the statistical significance in order to confirm the model consistency.

Findings: RCED-UNet3+ model had high generalization potential with a Dice score of 0.980 with consistent Intersection over Union (IoU) value of 0.961 on the Luna16 dataset. The findings validate the observation that segmentation accuracy of the model is also consistent despite changes in imaging parameters.

Conclusion: The research establishes that RCED-UNet3+ is a stable and efficient segmentation model in the detection of lung nodules. Its capability of sustaining high performance in datasets that contain mixed imaging properties of heterogeneous nature suggests its high potential integration into clinical workflows that seek to result in early and precise lung cancer diagnosis.