Hybrid cellular automata and U-Net framework for enhanced lung segmentation in chest radiographs

Accurate lung-field segmentation from chest radiographs is an important prerequisite for computer-aided diagnosis systems, yet it remains challenging due to anatomical variability, low contrast, overlapping thoracic structures, and boundary ambiguity. In this study, we propose a cellular automata (CA)-enhanced U-Net framework for lung-region segmentation in chest X-ray images. The proposed method first generates CA-derived anatomical priors through seed generation, iterative Moore-neighbourhood CA refinement, anatomical plausibility checking, region-of-interest-based rescue, soft-prior extraction, and boundary-prior extraction. The resulting soft CA prior and CA boundary prior are combined with the original grayscale chest X-ray image to construct a three-channel input representation. A trainable CA refinement layer further updates the CA-guided representation, and CA-gated decoder skip connections guide the segmentation network toward anatomically plausible lung regions. The predicted lung probability map is finally refined using validation-selected anatomical post-processing. The proposed framework was evaluated through multi-dataset experiments, k-fold cross-validation, cross-dataset evaluation, and comparison with a conventional U-Net and representative state-of-the-art lung segmentation methods. Experimental results demonstrate that the proposed CA-enhanced U-Net with post-processing provides consistent improvements over the baseline U-Net across overlap-based and pixel-level segmentation metrics. The incorporation of CA-derived soft and boundary priors improves spatial guidance, boundary localisation, and structural coherence of the predicted lung masks. Overall, the proposed hybrid rule-based CA and deep learning framework provides an effective and anatomically guided approach for lung-field segmentation in chest radiographs, with modest but consistent improvements over the baseline U-Net.