Page 52 - ARNM-1-2

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

In addition, the images produced often suffer from trained on diverse data, demonstrate better adaptability to
problems such as noise, low resolution, and low contrast, different modalities and scenarios.
which reduce image quality and diagnostic accuracy. In From January 2010 to March 2023, image fusion
clinical practice, other modality images, such as computed technology has made significant progress in NMMI, as shown
tomography (CT) and magnetic resonance imaging (MRI), in Figure 1. The number of publications on molecular imaging
are often used to supplement information (Figure S1). To and nuclear medicine image fusion has shown sustained
address the challenges mentioned above, image fusion growth in recent years (the number of papers was calculated
technology aims to merge multiple image data sources into using Web of Science from January 2010 to March 2023, with
a high-quality image, playing a significant role in disease keywords “medical image fusion,” “NMMI,” and “medical
diagnosis and treatment processes, including monitoring, multimodal image”). Several factors contribute to this growth:
characterization, and other processes . For instance, recent (i) Single-modality PET and SPECT images are insufficient to
[1]
studies have applied prostate-specific membrane antigen- meet the clinical diagnosis needs, necessitating the fusion of
PET/CT (PSMA PET/CT) image fusion technology to other modal images such as CT or MRI to provide additional
guide biopsies . Image fusion technology has become an information; (ii) the continuous improvement in computer
[2]
essential tool in the field of NMMI for improving image hardware performance enables the effective implementation
quality and information content .
[3]
of large-scale image fusion technology; and (iii) ongoing
Image fusion technology combines image information research and refinement of image processing algorithms have
from multiple imaging modalities to produce a fused image led to continuous enhancement in image fusion.
with improved diagnostic accuracy and visualization. Its This review provides an overview of the development
development can be traced back to the 1980s . Since then, of image fusion technology, introduces traditional image
[4]
image fusion technology has rapidly developed in various fusion techniques and deep learning-based image fusion
fields, particularly computer vision, medical imaging, methods, and finally discusses the challenges and future
and remote sensing imaging . In the early 1990s, image directions of image fusion technology in NMMI.
[5]
fusion technology began to be applied in medical imaging.
Early image fusion techniques relied on the manual 2. Traditional image fusion methods
alignment and superimposition of two images. Physicians
would visually compare different modalities of nuclear Traditional image fusion methods typically involve
medicine images and manually fuse them. This method is several steps. First, multiple modal images are acquired
complex, requires significant manual intervention, and is from different medical imaging devices. These images are
subjective, limiting the reproducibility and accuracy of the then preprocessed, including denoising, registration, and
resulting images. With the rapid development of computer calibration, to ensure spatial and intensity consistency.
technology and improvements in hardware equipment, Next, an appropriate fusion algorithm is selected to
automatic image registration and fusion algorithms have combine the multiple modal images. Finally, the quality and
emerged, providing strong support for the advancement effectiveness of the fusion result are evaluated. Traditional
of image fusion technology. In recent years, with the rapid image fusion methods can be mainly categorized into
development of deep learning technology, deep learning- five types: Spatial fusion, frequency fusion, decision-level
based image fusion methods have become a research fusion, hybrid fusion, and sparse representation fusion.
[6]
hotspot . These methods use neural network models
to learn the relationship between images and generate
higher-quality fused images . Traditional and deep
[7]
learning-based methods differ in several aspects. First, in
terms of methodology, traditional methods often rely on
predefined rules and handcrafted features for fusion, while
deep learning-based methods utilize neural networks to
learn and automatically extract features from input images.
Second, in feature representation, traditional methods
usually focus on low-level features, whereas deep learning-
based methods can capture complex and high-level features
by learning hierarchical representations. Finally, in regard
to adaptability, traditional methods often require manual
parameter tuning and adjustment for different imaging Figure 1. Number of published papers on image fusion in nuclear
modalities, whereas deep learning-based methods, once medicine molecular imaging from January 2010 to March 2023.

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