Application Research of Image Processing Based on Artificial

Intelligence in Medical Field

Xuanchen Dong

1,* a

, Liyuan Xiong

and Churui Zhang

Houston International Institute

Dalian Maritime University, Linghai Road, Ganjingzi District, China

Electronic Engineering,

Chengdu University of Information Technology, Xuefu Road, Shuangliu District, China

Electronic Information School

Wuhan University, Ba Yi Road, Wuchang District, China

Keywords: Artificial Intelligence, Medical Imaging, Deep Learning.

Abstract: This paper discusses in depth the contribution of image processing based on artificial intelligence to medicine.

Medical images include ultrasound images, X-rays and magnetic resonance images. However, interpreting

these images often relies on subjective judgment and professional experience, which is not only time-

consuming, but also prone to human error. Therefore, image processing plays a very important role in the

medical field. This paper summarizes the optimal AI solutions for ultrasonic, X-ray, and CT image processing

through a large literature review. In ultrasonic image processing, deep learning technology reduces the

reliance on subjective judgment and improves the accuracy of diagnosis by automatically learning image

features. In X-ray image analysis, models such as u-net+Attention have been developed to detect the high risk

factors for postoperative recurrence of lumbar disc herniation and improve the accuracy of image

interpretation. For CT images, deep learning models are applied to improve the accuracy of diagnostic results.

The paper also demonstrates the effectiveness of these AI techniques in different medical imaging applications

through experimental results. The application of this technology not only improves the accuracy and

efficiency of diagnosis, but also provides strong support for clinical treatment, which has important research

value and broad application prospects.

1 INTRODUCTION

With the popularity of the internet and mobile

devices, image data has exploded, which provides

rich materials for the training of deep learning. At the

same time, with the continuous progress of hardware

such as GPU, large-scale neural can be processed

efficiently. Lay the foundation for the application of

deep learning in the image recognition field. The

theory of deep learning continues to achieve

breakthroughs, especially the proposal and

development of convolutional neural Networks

(CNN), providing a powerful method for automatic

feature learning for image recognition. CNN can

automatically learn hierarchical feature

representations in images through multiple layers of

https://orcid.org/0009-0003-4886-4932

https://orcid.org/0009-0002-5472-7769

https://orcid.org/0009-0009-5234-8889

＊

Correspongding author

convolution and pooling operations, thereby better

capturing both local and global information in

images, significantly improving the accuracy of

image recognition.

Deep learning image recognition technology has

been applied in many industries, among which, it can

be used for the analysis and diagnosis of medical

images in the medical industry, such as the automatic

identification of X-ray, CT, MRI, and other images

and lesion detection. It can detect lesions quickly and

accurately, assist doctors in diagnosing diseases, and

improve the efficiency and accuracy of diagnosis,

which has great significance for the early detection

and treatment of diseases. This research focuses on

the three aspects of ultrasound, CT and X-ray

imaging. First of all, artificial intelligence plays a big

Dong, X., Xiong, L. and Zhang, C.

Application Research of Image Processing Based on Artiﬁcial Intelligence in Medical Field.

DOI: 10.5220/0013686600004670

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 2nd International Conference on Data Science and Engineering (ICDSE 2025), pages 271-276

ISBN: 978-989-758-765-8

271

role in processing ultrasound images.However there

is a problem that ultrasonic images require more

subjective judgment, which need to be solved.CADe

， CADx, Machine Learning (ML) and Deep

Learning(DL) are used for making the ultrasonic

image processing more accurate. Secondly, artificial

intelligence is used in X-ray images. In order to

explore the high risk factors of PETD surgery

recurrence, improve the clinical treatment effect of

lumbar disc herniation, through a large sample of

related risk factors after PETD operation long time-

course follow-up analysis, initially established the

model of the probability of PETD surgery recurrence

prediction and the full convolutional neural network

based on u-net + Attention. Finally, CT images,

previous studies used residual network (ResNet) for

image classification, and recurrent neural network

(LSTM) for the results correction. For the

segmentation task, a V-type network (VB-Net) is

used for model training to calculate the hematoma

volume.

Focusing on the above problems and challenges,

this paper analyzes and summarizes the research

progress and current situation of the utilization of

image recognition based on deep learning in the

medical field. This paper introduces the role of deep

learning in different medical images in detail, and

prospects that more medical images can be processed

by AI in the future.

2 METHOD

2.1 Ultrasonic Image

Artificial intelligence, especially machine learning

and deep learning technologies, plays an important

role in ultrasonic image processing These

technologies are able to automatically analyze and

interpret ultrasonic images, reducing the reliance on

manual interpretation. Computer aided detection（

CADe）and computer aided diagnosis（CADx）are

two key technologies in ultrasonic image processing.

CADe is used to automatically detect abnormal areas

in the image, helping doctors to find potential

problems in the image, especially in the initial

screening, can reduce the workload of manual

examination and improve the efficiency of diagnosis.

However, CADe is prone to false positives or false

negatives, leading to misdiagnosis or missed

diagnoses, and model performance is strongly

dependent on the quality and quantity of training data.

CADx is used for further analysis and diagnosis of

these abnormal areas. CADx not only provides test

results, but also can be used for diagnostic inference,

such as judging the benign and malignant nature of

the tumor, and the output is more granular, providing

information about the type and stage of illness. CADx

involves more reasoning and decision-making

processes, is computationally expensive, and the

black box nature of the model may lead to a lack of

trust among physicians when using it. By training a

large amount of ultrasonic image data, machine

learning and deep learning models are able to learn

the feature patterns in the image, thereby improving

the accuracy of image processing. These techniques

can be applied to image filtering, segmentation,

enhancement, etc., to improve image quality and

extract useful diagnostic information.

2.2 X-ray Image

VGG16 performs well in X-ray image classification,

mainly due to its deep network structure and 3×3

small convolutional kernel design, which can

effectively extract complex features from low to high

level, and is suitable for capturing subtle lesions such

as pneumonia and fractures in X-ray. Its transfer

learning ability is remarkable, and it can quickly

adapt to medical imaging tasks on small datasets

through pre-training weights and fine-tuning. In

addition, VGG16 has a simple structure, strong

interpretability, and supports feature visualization,

which is convenient for doctors to verify model

decisions. Despite the large number of parameters

and the high computational cost, these issues can be

mitigated through data augmentation, regularization,

and GPU acceleration. The maturity and stability of

VGG16 make it of great value in medical image

analysis, especially for the research and clinical pilot

of small and medium-scale datasets.

The u-net+Attention fully convolutional neural

network is used for automatic analysis and

interpretation of X-ray images. By introducing an

attention mechanism, the model is able to locate and

analyze key areas in the image more accurately,

thereby improving the accuracy of diagnosis. U-Net

combined with the Attention mechanism excels in

medical image segmentation, enabling accurate

segmentation of complex structures, such as tumors

or organs. With the introduction of the Attention

mechanism, U-Net can better focus on important

areas and improve the segmentation accuracy,

especially in complex or blurry image areas.

However, the U-Net structure requires large

computing resources, especially after the introduction

ICDSE 2025 - The International Conference on Data Science and Engineering

272

of Attention, and the training time is longer,

especially when training on large-scale datasets, it

may take a long time to converge.

Convolutional neural network (CNN) can

automatically learn the features in medical images,

especially the local structures in the images, such as

tumors and blood vessels, which are suitable for

classification, detection and segmentation tasks, and

have become the basic methods in medical image

analysis. However, CNNs require a large amount of

labeled data for training, and when the data is

insufficient, the performance of the model may be

unstable, and it is difficult to model the global

structure.

2.3 CT Image

Residual Network (ResNet) is a deep learning model

with powerful image classification capabilities. It is

capable of accurately classifying different types of

images by learning feature representations within the

images. In CT image processing, ResNet can be used

for identifying and analyzing lesions, organs, and

other structures. ResNet solves the gradient vanishing

problem in deep network training through residual

connections, allowing the network to be deeper and

more complex, making it very suitable for complex

medical image analysis, such as tumor detection and

lesion classification. However, the complexity of the

ResNet model is relatively high, increasing the

complexity of computation and training difficulty,

and it requires a larger and more diverse dataset.

Long Short-Term Memory (LSTM) is a special type

of recurrent neural network with the ability to learn

long-term dependencies. In CT image processing,

LSTM can be used to correct and optimize

preliminary classification results, improving the

accuracy of diagnosis. LSTM is particularly suitable

for processing time-series data and is applicable for

dynamic image analysis, such as sequences in

medical imaging, including dynamic scan data from

CT and MRI. LSTM can capture relevant information

over long periods, which helps to capture the

evolution of lesions over time. However, the structure

of LSTM is relatively complex, the training process

is more time-consuming. The structure of LSTM is

not adept at processing spatial features of images,

often combined with CNNs, which increases

computational complexity.

V-Net (VB-Net) is a deep learning model

specifically designed for medical image segmentation

tasks. It can automatically identify and segment

bleeding areas in CT images and automatically

calculate hematoma volume. This method can

provide doctors with more accurate and intuitive

diagnostic information. VB-Net uses variational

Bayesian methods for inference, which can provide

uncertainty estimates for medical image diagnosis,

helping to handle complex and ambiguous medical

images. VB-Net is robust to noise and missing data,

but Bayesian inference methods usually involve

complex probabilistic calculations, and the training

and inference processes are relatively slow, and

implementation and debugging are more complex.

3 APPLICATION RESEARCH

Artificial intelligence plays a big role in processing

ultrasound images. Ultrasound image processing is

more difficult than MRI because ultrasound images

require more subjective judgment. In order to solve

this problem, CADe and CADx have become the

main solutions, but CADe and CADx still have a

large degree of limitations in the collection and

processing of data. Machine learning (ML) and deep

learning (DL) can make ultrasonic image processing

more accurate. Machine learning can only identify

basic cases from the data set, so relying solely on

machine learning to analyze disease categories based

on ultrasound images becomes more difficult because

many diseases vary too much for the data set to cover

all cases. Lee YH and Shin Y of Yonsei University

found that DL was much more accurate than ML in

classifying skeletal muscle diseases (Shin, Yang, Lee,

2021). This is because deep learning can imitate the

neural network of human brain and use a large

number of labeled data for learning training (LeCun,

2015). In order to identify Duchenne muscular

dystrophy (DMD), echo intensity and muscle

thickness were quantified to distinguish echo-

enhanced myopathy muscles, and subsequently ML

and DL were used to compare ultrasound images of

patients with normal muscle images and

automatically partition the parts containing body

myositis, polymyositis, dermatomyositis, and normal

manifestations. In this process DL is more accurate

than ML regardless of the type of skeletal muscle

disease. It can be seen that DL technology has

become the core technology of ultrasonic image

processing.

Cal Shaohui, Lin Qiaojuan et al. introduced AI

into the diagnosis of pulmonary nodules. The

thickness of 64 row spiral CT and 64 row CT layer of

Somatom Definition AS model was set to 1 mm,

Application Research of Image Processing Based on Artiﬁcial Intelligence in Medical Field

273

followed by interval-free reconstruction, the window

width was set to 1 450 HU and the window position

was set to-831.1HU to ensure accurate identification

of lesions.Related data (such as communication data,

medical digital imaging, etc.) after import relevant

imaging examination, AI pulmonary nodule

diagnosis system, will identify and check two stages,

then use computer-aided quantitative parameter

system data preprocessing, single box detector

simulation training, nodule measurement operation,

automatically check the nodule edge, this series of

operation do the three-dimensional state of nodule

long diameter, short diameter, volume, maximum

cross-sectional area and other parameters. However,

the study of Li Juan, Tang Xiangyu et al. used AI

analysis for intracranial hematoma, using two

datasets (Li, Tang, 2021). Data set 1 contained 9594

plain scan images of craniocerebral CT, of which 223

patients with positive intracranial hemorrhage were

used as the test set, and the rest were used as the

training set. Dataset 2 contains 819 CT images of

bleeding foci that have been manually delineated, of

which 74 were used as a test set to verify the

consistency between algorithm segmentation and

manual segmentation. Data input of the CT images

was first performed, and all of the CT images were

performed in a standard DICOM format. Data

preprocessing included image correction, skull

removal, and grayscale normalization. In the cross-

axis data, the position of the two endpoints in the

midline brain is detected based on deep learning, and

the CT image of the cross-axis is rotated, which

automatically positions the brain position. Then, the

brain tissue area was automatically segmented based

on deep learning, and the interference information

including the skull and other images in the image, was

automatically removed. After the gray scale

normalization to [-1,1], the calling residual network

(ResNet) of the image was classified according to five

bleeding types and a total of six labels. For each layer

of classification results, call circulating neural

network (long short-term memory network, LSTM)

results correction.For bleeding focus segmentation

task, after image pretreatment, call V network (VB-

Net) model training, which through the voxel

statistics and spacing conversion, the introduction of

the algorithm and the model can automatically get

each bleeding statistics to calculate the hematoma

volume, is a big improvement.

Jinxiu Cai et al. studied a deep learning-based

chest X-ray (CXR) image classification model(Cai,

2022), using the Vgg16 network to classify different

types of chest X-rays, and successfully distinguished

between adult anteroposterior chest x-rays, lateral x-

rays, bedside x-rays, and infant x-rays. The model

showed high accuracy (94%~100%) in the test set and

external validation, and was able to automatically

screen out qualified images for subsequent disease

diagnosis. The characteristics of this research are that

the automation level of image classification is

improved through deep learning models, the errors of

manual operation are reduced, and the work

efficiency of the imaging department is improved. In

the future, the model is expected to be further

optimized and extended to more image quality

evaluation scenarios.

Ramadhan Hardani Putra et al. reviewed the

application of artificial intelligence in digital dental

radiology, covering multiple aspects such as caries

diagnosis, periodontal bone loss analysis, cyst and

tumor classification, etc (Ramadhan, 2022). The

study has shown that deep learning (DL) and

convolutional neural networks (CNNs) excel in

dental image analysis, automatically identifying

complex image patterns and providing quantitative

analysis. The research is characterized by the fact that

it demonstrates the potential of deep learning for a

wide range of applications in dental imaging,

especially in terms of automated diagnosis and image

quality enhancement. In the future, with the

expansion of datasets and the improvement of

algorithms, deep learning is expected to play a greater

role in dental clinical practice.

Finally, Xian Chang et al studied the identification

of key parameters of lumbar X-ray based on deep

learning model, and constructed a fully convolutional

neural network based on U-net+Attention, which was

used to automatically measure the intervertebral

space height index, lumbar spine motion angle and

segmental mobility (Xian, 2024) . The average IOU

of the model on the test set reached 0.940 and the Dice

coefficient was 0.980, showing high segmentation

accuracy. The characteristics of this study are that the

automatic measurement of lumbar spine imaging

parameters is realized through deep learning

technology, which reduces the error and workload of

manual measurement, and improves the accuracy of

clinical decision-making. In the future, the model is

expected to be further optimized and applied to more

spine image analysis scenarios. In summary, the

contribution of deep learning in the field of X-ray is

mainly reflected in automation, high precision and

high efficiency. Through different deep learning

models, researchers have successfully solved the

problems of medical image classification,

segmentation, and parameter measurement, and

significantly improved the efficiency and accuracy of

ICDSE 2025 - The International Conference on Data Science and Engineering

274

image analysis. In the future, with the continuous

advancement of technology and the expansion of data

sets, the application of deep learning in the field of

medical imaging will be more extensive and deeper,

which is expected to provide stronger support for

clinical diagnosis and treatment.

4 CONCLUSION

This paper discusses the application and effect of

artificial intelligence in ultrasonic image, CT image

and X-ray image processing. The results show that

artificial intelligence technology, especially deep

learning methods, has shown significant advantages

and great potential in the field of medical imaging

diagnosis.

In ultrasonic image processing, deep learning

overcomes the limitations of traditional CAD systems

by imitating the human brain's neural network and

using lots of labeled data for training. This greatly

improves the accuracy of skeletal muscle disease

classification. This technological advance not only

improves the reliability of diagnosis, but also

provides stronger support for the early detection and

treatment of related diseases.

In CT imaging diagnosis, the application of AI

technology has realized the automatic detection and

accurate parameter determination of lung nodules and

the efficient analysis of intracranial hematoma. The

deep learning algorithm lets the system automatically

identify and classify bleeding types, calculate

hematoma volume, and provide quick, accurate

diagnosis for clinicians. This helps improve patient

outcomes by enabling timely treatment.

For X-ray images, the deep learning-based model

performed well in the identification of key parameters

in the lumbar spine and the analysis of chest X-rays.

The model of automatic measurement of lumbar

parameters provides a powerful tool for predicting the

risk of surgical recurrence, reducing human error and

improving work efficiency. At the same time, the

binary diagnosis AI model of chest X-ray film has

reached a high accuracy in the diagnosis of "abnormal

finding" and "no abnormal finding", which

effectively improves the daily work efficiency of the

medical imaging department.

To sum up, the application of artificial

intelligence technology in medical image diagnosis

not only improves the accuracy and efficiency of

diagnosis, but also provides more powerful support

for clinical treatment. With the continuous progress

of technology and the accumulation of data, the

application of artificial intelligence in the field of

medical imaging will be more extensive and in-depth,

which is expected to further promote the development

of medical imaging diagnostic technology and

improve the treatment effect and quality of life of

patients. Future research should continue to explore

more efficient and accurate algorithms, optimize

model performance, and strengthen multi-center

cooperation to promote the widespread application

and standardization of AI technology in medical

image diagnosis.

AUTHORS CONTRIBUTION

All the authors contributed equally and their names

were listed in alphabetical order.

REFERENCES

Cai, J., Ren, X., Sun, Z., et al. (2022). Preliminary study of

artificial intelligence model of chest X-ray image

properties based on deep learning. The Journal of

Practical Radiology, 38(7), 1184-1188.

Cai, S., Lin, Q., Yang, S., Chen, S., & Su, L. (2024).

Clinical value of CT and AI diagnostic system for

diagnosis and differential classification of pulmonary

nodules. Medical Equipment, 37(5), 30-34.

Chang, X., Zhang, B., Wang, X., et al. (2024). Application

of the key parameter identification based on deep

learning model. Journal of Orthopaedic Clinical and

Research, 9(1), 31-36.

Lecun, Y., Bengio, Y., & Hinton, G. (2015). Deep learning.

Nature, 521(7553), 436-444.

Li, J., Tang, X., Shen, Y., et al. (2021). Research on the type

identification and hematoma segmentation consistency

of intracranial hemorrhage by deep learning algorithm

based on convolutional neural network. Radiology

Practice, 36(1), 7-12.

Nair, M.M., Kumari, S., Tyagi, A.K., & Sravanthi, K.

(2021). Deep Learning for Medical Image Recognition:

Open Issues and a Way to Forward. In D. Goyal, A.K.

Gupta, V. Piuri, M. Ganzha, & M. Paprzycki (Eds.),

Proceedings of the Second International Conference on

Information Management and Machine Intelligence.

Lecture Notes in Networks and Systems, vol. 166.

Springer, Singapore.

Putra, R.H., Doi, C., Yoda, N., Astuti, E.R., & Sasaki, K.

(2022). Current applications and development of

artificial intelligence for digital dental radiography.

Dento-Maxillo-Facial Radiology, 51(1), 20210197.

Razzak, M.I., Naz, S., & Zaib, A. (2018). Deep Learning

for Medical Image Processing: Overview, Challenges

and the Future. In N. Dey, A. Ashour, & S. Borra (Eds.),

Classification in BioApps. Lecture Notes in

Computational Vision and Biomechanics, vol. 26.

Springer, Cham.

Application Research of Image Processing Based on Artiﬁcial Intelligence in Medical Field

275

Yang, Y., Lee, Y.H., & Kim, S. (2021). Artificial

intelligence in musculoskeletal ultrasound imaging.

Ultrasonography (Seoul, Korea), 40(1), 30-44.

Zhang, Z., Wang, J., & Liu, T. (2020). Deep learning in

medical imaging: A review. Journal of Healthcare

Engineering, 2020, 8907850

ICDSE 2025 - The International Conference on Data Science and Engineering

276