Page 54 - 2025S

UEC Int’l Mini-Conference No.54 47

Figure 10: Image extracted from the DICOM image
by zooming in on the area of interest
(a) Original grayscale img. (b) Reconstructed image

Figure 11: (a) Original grayscale image, (b) recon-
1: for each block B i,j in I do structed image using the VDSR network.
′
2: for each pixel p r,c ∈ B i,j (except p 1,1 ) do
3: b ← LSB(p r,c )
4: B ext ← B ext ∥ b stantial improvement, obtaining a PSNR of 67.50
5: end for and an SSIM of 0.9995. This not only represents
6: end for a significant enhancement in reconstruction fidelity,
7: D ← ChaoticDecryption(B ext ) but also reflects better computational efficiency. The
After the watermark extraction, the embedded new approach successfully restores fine image de-
halftone image is first localized within the artifi- tails with higher visual accuracy, making it more
cially padded region. This step is crucial for accu- suitable for clinical scenarios where image quality
rately isolating the hidden image from the surround- is critical for patient identification and diagnostic in-
ing padding. Once localized, the artificial border is tegrity.
removed, and the halftone image undergoes a back-
ground separation process that distinguishes the sub- 4 Conclusions
ject from the structured vertical-line background.
This background region contains the embedded This study presents a novel reversible data hiding
binary message, which is then decoded. In our ex- scheme for DICOM medical images that combines
periments, a total of 51,652 bits (equivalent to 6,456 halftone watermarking with chaotic encryption and
characters) were successfully extracted. This re- deep learning-based reconstruction. The proposed
covered message corresponds to the patient’s con- method addresses three critical challenges in med-
fidential information. At the end of this stage, the ical image security: (1) high-capacity embedding
halftone image remains intact and isolated—ready of patient identification data, (2) tamper detection
for subsequent grayscale reconstruction in the next through fragile watermarking, and (3) perfect re-
processing stage. versibility to maintain diagnostic quality.
This approach provides radiologists and health-
3.4 Reconstruction stage care systems with a practical tool for patient-data
linkage while meeting the strict reversibility require-
In the final stage of the pipeline, we implement ments of medical imaging. The combination of high
a new alternative for grayscale reconstruction of
the halftone image. In previous iterations of our capacity, cryptographic security, and diagnostic-
method, we employ the Very Deep Super-Resolution grade reconstruction makes it particularly suitable
(VDSR) neural network for this task. While VDSR for teleradiology and forensic applications where
produces acceptable results—yielding a PSNR of image authenticity is paramount.
26.0272 and an SSIM of 0.8851—its performance is
limited both in terms of perceptual quality and com-
putational efficiency [13].
With our new implementation, we achieve a sub-