Texture-aware reversible information embedding in medical images using attention-driven dual-branch deep prediction

This work proposes a reversible information embedding (RIE) technique for grayscale medical images, aiming to provide secure and fully reversible data concealment without compromising diagnostic integrity. The proposed approach employs a CBAM-assisted dual-branch neural network predictor (DBNNP) to enhance pixel prediction accuracy through independent learning of shared and local features. Each medical image is partitioned into four interleaved sub-images using a checkerboard pattern, where two sub-images act as predictors and the remaining two serve as targets for data embedding. Texture-driven clustering based on Local Standard Deviation (LSD) and Local Weighted Brightness (LWB) is used to generate region masks that distinguish non-regions of interest (NROI) from texture-varying regions of interest (ROI). These masks guide the dual-branch network to enable region-aware prediction through improved spatial and channel attention mechanisms. Following prediction, a selective dual-bit prediction error expansion scheme is applied exclusively to zero-prediction-error pixels, ensuring full reversibility while increasing embedding capacity. The proposed framework integrates texture-aware preprocessing, attention-guided deep prediction, and controlled error expansion to form a robust and clinically viable RIE solution. Experimental evaluations on medical X-ray datasets demonstrate that the method preserves visual quality while enabling accurate data extraction and perfect image recovery, making it suitable for secure medical image transmission, archival, and telemedicine applications.