Page 54 - ARNM-1-2

Advances in Radiotherapy
& Nuclear Medicine Image fusion’s significance in medicine

combined the IHS model with Log-Gabor transform and distribution in each input image. Subsequently, it employs
proposed a new method for fusing MRI and PET, using a set of decision rules, such as maximum a posteriori,
IHS to decompose PET images into three basic features: maximum likelihood, and expectation, to determine the
Hue, saturation, and intensity. This approach effectively final value of each pixel. The fusion rules of this approach
preserves the structure and details of the source image and rely on statistical analysis and decision theory, enabling it
reduces color distortion. to fully utilize the information from each input image. As
a result, decision-level image fusion improves the quality
2.2. Frequency fusion and robustness of the fusion results .
[17]
Frequency image fusion methods are based on the A non-parametric Bayesian technique is considered to
principles of Fourier transform. These methods involve learn dictionaries and mappings for two feature spaces .
[18]
transforming images from the spatial domain to the In the proposed method, dictionaries for two feature spaces
frequency domain. The fusion algorithm is then applied to and the mapping between them are adaptively obtained
the transformed images, and the inverse Fourier transform from input PET and MR images. This algorithm not only
is performed to obtain the synthesized fusion image. learns the dictionaries for each feature space separately but
Frequency fusion techniques can be further classified into also considers the relation between two feature spaces. This
pyramid and transform-based methods . Compared to relation helped establish a proper connection between two
[13]
spatial fusion methods, the frequency fusion method is different input images.
more complex, but it offers the advantage of reducing the
distortion of the fusion image. 2.4. Sparse representation fusion
Liu et al. proposed a medical image fusion algorithm Sparse representation fusion combines sparse
[14]
based on multiwavelet transform for PET/CT fusion. The representation theory and image fusion, making it a
experimental results demonstrated that the fusion image highly effective medical image fusion method. The
integrates information from the source images, adding method operates on the assumption that an image can
more details and texture information and ultimately be represented linearly using a small number of atoms
achieving a good fusion result. Haribabu et al. proposed (such as the basis vectors in a dictionary). Initially, the
[15]
a new approach for PET-MRI image fusion using wavelet input images are decomposed into a set of coefficients,
and spatial frequency methods. This algorithm addresses representing linear combinations of the basis vectors in the
the issues of image imbalance and blurred phenomena dictionary. Subsequently, these coefficients are combined
often encountered infusion images, improving clarity to generate fusion coefficients, which produce the fused
and providing more reference information for medical image. By utilizing only a small subset of the basis vectors,
professionals. this method excels at preserving the features of the original
images while reducing the influence of image distortion
In addition to transform-based methods, the pyramid and noise .
[19]
technique is also commonly used in image fusion. This
technique decomposes the original image into multiple Zong and Qiu [20] developed a new fusion scheme
scales, fuses the corresponding scales, and then reconstructs for medical images based on a sparse representation of
to obtain the fused image. During this process, Gaussian classified image patches. To achieve this, image patches
pyramid (GP) and Laplacian pyramid (LP) are typically are first classified according to their geometrical gradient
employed for image decomposition and reconstruction. direction. Subsequently, multi-class dictionaries are
Sahu et al. [16] proposed an algorithm that utilizes LP with trained within each class using the online dictionary
discrete cosine transformation (DCT). The LP decomposes learning (ODL) algorithm.
the input image into different low-pass images, creating a 2.5. Hybrid fusion
pyramid-like structure. As the pyramid levels increase, the
quality of the fusion image improves, thus enhancing the Considering the limitations of traditional image fusion
edges and information. In both quantitative and qualitative methods, hybrid methods have gained popularity.
analyses, this method outperformed Daubechies complex Hybrid methods involve combining two or more fusion
wavelet transform (DCxWT), producing excellent fusion techniques, such as spatial fusion and frequency fusion, to
results. utilize information from different domains and improve
the image quality and lesion detection capabilities.
2.3. Decision-level fusion
[21]
Zhu et al. decomposed source multi-modality images
Decision-level image fusion treats the value of each into cartoon and texture components. They proposed a
pixel as a random variable and calculates its probability proper spatial-based method for preserving morphological

Volume 1 Issue 2 (2023) 4 https://doi.org/10.36922/arnm.0870