Page 50 - 2025S

UEC Int’l Mini-Conference No.54 43

ical operations, such as bitwise NOT, guided by the
chaotic state, ensuring high entropy in the encrypted
output.

To ensure authenticity and enable tamper detec-
tion, a checksum is computed for each image block
by nullifying the two least significant bits (LSBs) of (a) Original image (b) Noisy image (c) Denoised img.
every pixel and calculating the sum of the modified
block. This value is converted into binary form, and Figure 3: Results obtained using the Denoising Con-
volutional Neural Network: (a) original image, (b)
a checksum vector is generated using XOR opera- noisy image, and (c) denoised image.
tions on the bit pairs. During the embedding stage,
the encrypted EPR and watermark are placed into
the intermediate significant bits (ISBs), while check- 3 Proposed Method
sum bits are embedded into specific rows of the im-
age block. This design allows the system to detect The general diagram of the proposed method is
localized tampering while preserving reversibility.
shown in Figure 4.
Additionally, the method improves embedding ca-
pacity by first duplicating the size of the image prior
to data hiding. This upscaling step increases the
number of available pixels for secure data insertion
without distorting the image’s visual structure. The
combination of chaotic encryption, checksum-based
authentication, and spatial expansion forms a com-
prehensive framework for secure and reversible data
hiding in clinical imaging applications [10].

Figure 4: General diagram of the proposed method

3.1 Insertion stage

2.4 DnCNN Network for Halftone Image The initial step involves processing the patient’s
Enhancement photograph according to the specifications defined
in the system architecture. To this end, the halfton-
ing technique is applied—specifically, Jarvis error-
The DnCNN (Denoising Convolutional Neural Net- diffusion method [12]—due to its effectiveness in
work) is a deep learning architecture designed to re- preserving fine details during the binarization of the
move noise from images. Implemented in MATLAB grayscale image.
through the denoisingNetwork(”DnCNN”) function, In contrast to previous approaches that relied
this network has demonstrated strong performance on manually generating artificial contours around
in reducing Gaussian noise in grayscale images. specific pixels to expand the data embedding ca-
pacity, our current method eliminates the need for
DnCNN operates by learning a residual represen- such vector-based outlining. Instead, we introduce
tation of the noise present in the input image, allow- a segmentation-based preprocessing stage that iso-
ing for accurate restoration of the original content. lates the subject in the image using intensity-based
This capability is particularly beneficial in halftone thresholding and morphological operations.
image processing, where denoising can improve vi- Once the object of interest—typically the pa-
sual quality without compromising essential struc- tient’s silhouette—is segmented, we enhance the
tural details [11]. surrounding area with an artificial background com-