Challenges and Opportunities in Integrating

Machine Learning with Medical Imaging: A

Comprehensive Review
Manish Chaudhary Himanshu Agrawal
Computer Science and Engineering & Information Technology Computer Science and Engineering & Information Technology
2023 Second International Conference on Informatics (ICI) | 979-8-3503-4383-0/23/$31.00 ©2023 IEEE | DOI: 10.1109/ICI60088.2023.10421032

Jaypee Institute of Information Technology Jaypee Institute of Information Technology

Noida, U.P. Noida, U.P.
[email protected] [email protected]

Abstract- In recent years, there has been a significant from electronic health records [5].
increase in the usage of electronic medical records and The ascendancy of machine learning over conventional
diagnostic imaging, which coincides with the enormous biostatistical methods has become evident due to their
success that machine learning algorithms have had in aptitude for managing and comprehending vast, intricate
performing image identification jobs in recent years. This healthcare data sets [6]. Recent accomplishments in
article is an introduction to machine learning techniques as
machine learning encompass diverse tasks such as
they are applied to medical image processing, with a specific
focus on convolution neural networks as well as the clinical identifying body organs from medical images [7],
aspects of the issue. In this era of medical big data, one of the categorizing interstitial lung diseases [8], reconstructing
benefits of machine learning is that critical hierarchal medical images [9,10], and segmenting brain tumors. It
linkages within the data may be discovered algorithmically attests to its potency in delivering remarkable outcomes
rather than having to painstakingly hand-craft features. It across various medical applications.
saves a significant lot of time and effort. We cover major
research areas and applications in medical image
classification, localization, detection, segmentation, and • Device Mimics Cognitive
registration sectors. In conclusion, we will discuss research AI Functions
• Since 1950s
challenges, emerging trends, and probable future
possibilities.
Keywords- ML, AI, DL, medical imaging, healthcare Machine • Algorithms that inprove as
they are exposed to more
datat

I. INTRODUCTION
Learning • Since 1980s

Electronic medical data and diagnostic imaging have Deep • Artificial neural networks
structured in multiplie layer
to decode imaging raw data
grown a lot in recent years. It goes along with machine learning • Since 2010s
learning algorithms' considerable success in image
identification jobs in recent years. This article gives an
overview of how machine learning methods are used to
process medical images, focusing on convolutional neural
networks and the clinical aspects of the topic [1]. In this Fig. 1. AI generation
age of "big data" in medicine, one of the benefits of AI and ML rapidly reshape the scientific field,
machine learning is that it can find critical hierarchical significantly impacting various domains, including
links in the data without having to make features by hand. medicine. AI is making software or hardware that can
It gets rid of a large amount of work. This article discusses think and act like humans. Machine learning, or ML for
significant study areas and applications in medical image short, is a subfield of artificial intelligence that teaches
classification, localization, detection,[2] segmentation, computers to learn from data to improve their decision-
and registration. In the end, we will talk about the making, predictions or classifications, either with or
problems that research meets, some new trends, and some without human supervision. The rise of high-performance
likely possibilities for the future [3]. The predictive computers has greatly facilitated recent progress in these
models in machine learning evolve and make more fields.
precise decisions as they accumulate larger volumes of In the medical field, areas that have been digitized, such
data. In healthcare, patient information stored in as medical imaging, are at the forefront of adopting AI
electronic health records can be effectively harnessed and ML technologies. The entire medical imaging
using machine learning techniques to extract relevant process, from image acquisition to interpretation,
insights [4]. Disease diagnosis benefits from machine reporting, and result communication, occurs within the
learning algorithms, as they scrutinize data to anticipate digital realm, making it well-suited for integration with
the root causes of illnesses based on variables sourced AI and ML. Given its substantial role in many medical
departments, cancer imaging is remarkably poised for
979-8-3503-4383-0/23/$31.00 ©2023 IEEE early adoption by radiologists, who serve as primary users

Authorized licensed use limited to: VIT University. Downloaded on May 13,2025 at 10:21:32 UTC from IEEE Xplore. Restrictions apply.
of these technologies. It is especially true because tasks in development of public health initiatives. The paper
this area can be repetitive (like sifting through numerous presents a comprehensive overview of the application of
regular studies to detect abnormalities during cancer machine learning techniques, specifically delving into
screening), monotonous (for instance, measuring tumor deep learning methods like CNNs and generative
size over time), and labor-intensive (such as manually adversarial networks (GANs), within the realm of medical
outlining tumors for disease segmentation). Existing imaging. It underscores the increasing significance of AI
commercial products are already focused on cancer and machine learning (ML) in the field of medicine. Here
imaging, aiming to enhance workflow efficiency, is a summary of the key contributions:
decrease errors, and improve diagnostic accuracy. Clinical Applications: The paper examines various
While numerous technological solutions are being clinical applications of machine learning in the processing
developed in separation, they might face challenges of medical images, encompassing tasks such as image
becoming part of routine clinical practices. It could be due classification, localization, detection, segmentation, and
to a need for more opportunities for collaboration among registration. It illustrates how these applications enhance
clinicians, radiologists, scientists, and other experts. the precision and efficiency of medical diagnoses and
These stakeholders must work together to comprehend treatment procedures.
clinical and data science aspects and identify Comparison of Machine Learning Algorithms: The paper
requirements, risks, opportunities, and obstacles in conducts a comparative analysis of diverse machine
developing, testing, validating, and adopting such tools. It learning algorithms, outlining their respective advantages
necessitates fostering multidisciplinary ecosystems with and drawbacks when applied to medical image analysis.
appropriate commercial partners to drive innovation and This comparative analysis aids researchers and
progress [11]. practitioners in selecting the most suitable algorithms for
Overview of Machine Learning: The process of training specific tasks.
algorithms with data to make educated guesses or carry Discussion of Challenges: The paper conducts a
out actions without being explicitly programmed is called comparative analysis of diverse machine learning
machine learning (ML), a subfield of artificial algorithms, outlining their respective advantages and
intelligence. The potential for machine learning to alter drawbacks when applied to medical image analysis. This
how diseases are diagnosed, treated, and prevented is comparative analysis aids researchers and practitioners in
illustrated in Figure 1. There are many ways machine selecting the most suitable algorithms for specific tasks.
learning could be helpful in the medical area, including: Focus on GANs: The paper emphasizes the role of GANs
in generating synthetic medical images for data
augmentation and other applications. It discusses the
challenges and opportunities associated with GANs in
medical imaging. The paper discusses the potential
impact of machine learning on healthcare, including early
disease detection, personalized treatment planning, and
public health programs. It underscores the transformative
potential of these technologies.
The remaining paper is organized as follows. Section II
Fig. 2. Concept of machine learning in medical imagining [10] discusses the two primary subfields of machine learning.
Section III explores the types of health datasets used in
Predictive Analytics: Data-driven algorithms as shown in medical imaging and their significance. Section IV
Figure 2, that learn from machine information can analyze describes the feature extraction process and its
the information included in patients' electronic health importance in machine learning. Section V explains the
records, insurance claims, and other sources to determine state-of-the-art ML methods for Medical Image Analysis.
the likelihood of specific health occurrences occurring. Section VI discusses the challenges faced in integrating
Some of these health events involve returning to the machine learning into medical practices. Finally, section
hospital or getting a new chronic sickness. Because of VII summarizes the key points discussed in the paper.
this, it is now much easier for medical professionals to
spot people who pose a severe risk and take steps to stop II. REVIEW OF MACHINE LEARNING
bad things from happening. Medical pictures like CT
scans and X-rays can be used to train machine learning As shown in Figure 3, there are two primary subfields that
algorithms, which allows them to help doctors diagnose makeup machine learning: supervised learning and
illnesses and figure out the best treatment for each patient. unsupervised learning. Both of these subfields can be
Learning machines can guess which treatments will most further subdivided. In supervised machine learning,
likely work for each patient based on their genetic algorithms are trained to generate predictions about future
makeup and medical histories. When ML methods are outputs using data already established as both an input
added to clinical decision support systems, they can help and an output. The goal of unsupervised machine
doctors and nurses make better decisions about caring for learning, on the other hand, is to unearth previously
their patients. ML can be used to look at data about a hidden patterns or structures that are already present in the
community to find patterns and trends. This information input data. Supervised machine learning effectively
offers valuable insights that can contribute to the conducts classification and regression tasks, whilst

Authorized licensed use limited to: VIT University. Downloaded on May 13,2025 at 10:21:32 UTC from IEEE Xplore. Restrictions apply.
 unsupervised machine learning is utilized for data natural language processing to learn more about patients'
clustering activities [12]. Supervised machine learning is health and develop focused ways to improve patient care.
also effective at doing tasks involving clustering.
Supervised machine learning classification methods IV. FEATURE EXTRACTIONS
encompass algorithms designed to predict categorical
outcomes, called classifications. Unlike unsupervised Choosing the most essential parts of a data set is an
learning, these techniques require labelled data where the integral part of machine learning and a vital part of the
relationships between inputs and outputs are already process as a whole. This process is called "feature
known. The data is divided into training and testing extraction." As part of this process, raw data are turned
subsets [13]. Classification algorithms within supervised into features that can be used to find trends and are easy
learning make predictions by assigning input data to to remember. Because of this change, finding trends in the
specific categories. Typical applications of classical original data will likely be more challenging than in the
supervised machine learning involve predicting heart extracted features. The main goal of feature extraction is
attacks, processing medical images, and recognizing determining which characteristics or features in the raw
speech [14]. data are the most important. After that, these will be used
as inputs for an algorithm that uses machine learning to
do a specific job [15]. LDA stands for linear discriminant
analysis, whereas PCA is an abbreviation for principal
component analysis. Some examples of methods that
could be used to collect features include autoencoders,
filter methods, wrapping methods, and the t-distributed
stochastic neighbor embedding, also known as t-SNE.
There are several more ways to approach this.

PCA: Principal Component Analysis is a method used in

data analysis and machine learning to lower the number
Fig. 3. Machine learning has a general structure of dimensions in a data set. PCA is a term that stands for
this method. It works as a linear transformation, and the
These algorithms construct classification models using goal is to find trends in high-dimensional data by mapping
training data. These models can then be employed to those dimensions to a lower-dimensional space. PCA was
classify other unlabeled data. A target for classification is developed to detect the most significant differences in the
included in the dataset used for training, and it is data while reducing noise and duplication [16].
represented as an output variable. Every algorithm is
designed first to recognize distinct patterns in the training LDA: Linear Discriminant Analysis, commonly referred
data and then to apply those patterns, after being refined, to as LDA, represents a technique employed to reduce the
to fresh test data for classification. Several well-known dimensionality of data under expert guidance. This
examples of machine learning approaches for supervised approach finds widespread application in the fields of
classification include the decision tree, the support vector machine learning, pattern recognition, and statistical
machine, the naive Bays approach, the K-nearest analysis. The fundamental goal of LDA is to condense
neighbor methodology, and the neural network. These are extensive datasets into a more compact form while
only a few examples. Figure 3 presents an overarching simultaneously enhancing the ability to distinguish
diagram of the machine learning infrastructure. between various classes within the data.
III. HEALTH DATASETS Autoencoders: A subset within the realm of artificial
neural networks is referred to as autoencoders. Unlike
Vast information about people's health is put together in other networks, autoencoders operate without the
healthcare records. These data sets include a wide range requirement for constant activity monitoring. They
of information, from a person's medical history and can be used to reduce the number of dimensions, extract
diagnostic test results to how they use their medicines and features, and learn how to describe data. They are also
their demographic information. They are used in clinical called "autoencoders." These tools are made up of two
studies, monitoring public health, and projects that try to parts: encoders and decoders. Together, they compress
improve things. Healthcare datasets include electronic incoming data and then rebuild it, losing less information.
health records (EHRs), computer files with medical Autoencoders can be used for denoising, finding outliers,
information about patients, and claims datasets that show and pre-training for complex neural networks without
what services were provided and how much they cost. being watched.
Also, disease records have information about people with
specific diseases or conditions, and clinical trial files have V. STATE-OF-THE-ART ML
information about the people who participated in clinical METHODS FOR MEDICAL IMAGE
trials, their treatments, and how the trials went. These ANALYSIS
health records are hard to understand because they are
long and hard to understand. However, scholars can use In the past decade, there has been a significant rise in
advanced analytic methods like machine learning and research and use of AI techniques, especially ML and DL
Authorized licensed use limited to: VIT University. Downloaded on May 13,2025 at 10:21:32 UTC from IEEE Xplore. Restrictions apply.
methods, in the medical area. These methods have been detection, localization, segmentation, and image-based
extensively used and have shown promise in several prediction.
medical settings. Over the last ten years, the focus has RFs have demonstrated strong performance in various
shifted from traditional ML methods like SVM and contexts compared to classical ML methods. For instance,
Random Forests to more advanced DL methods like in predicting outcomes and toxicity in
CNN. This change could react to the fact that medical (chemo)radiotherapy, when compared to other
data is getting more complicated and needs more classification techniques such as decision trees, neural
complex algorithms to handle it well. CNNs and other networks, and support vector machines, RFs were able to
Deep Learning methods have become popular because attain a greater level of discriminative performance in
they can automatically learn complex features from raw specific datasets and more. However, studies have
data. CNNs are especially good at jobs that involve highlighted that the performance of RFs can vary based
images, and they have been used a lot in medical image on the particular application and datasets, making it
analysis to do things like segmenting images, finding challenging to declare them superior to other ML
objects, and classifying diseases. classifiers definitively.
Expansion to Other DL Techniques: Alongside CNNs, Convolutional neural networks (CNNs)
other Deep Learning techniques like GANs and Convolution Neural Networks (CNNs) have become a
reinforcement learning algorithms have also gained dominant force in modern medical image processing due
traction in medical applications, especially since 2018. to their ability to harness the spatial organization of image
data, mimicking aspects of the human visual system.
CNNs are particularly adept at discerning hierarchical
data representations, making them highly effective for
tasks involving medical images. Traditionally, CNNs
consist of stacked convolution and down sampling layers,
which are interconnected and contribute to the outcome.
The convolutional layers progressively extract features
from bare edges to more complex hierarchical structures
like organs by applying convolution kernels across the
Fig. 4. The linear support vector machine [25]
image. Down sampling layers, often through max-pooling
operations, reduce the feature map size while retaining
GANs are known for their ability to generate new data
essential features. This trade-off between resolution and
samples that closely resemble the training data
feature richness allows more convolution filters to operate
distribution, which has found applications in generating
on smaller feature maps, optimizing memory usage. Fully
synthetic medical images for research and training
connected layers generate the final output. FCNs were
purposes. Nonlinear Class Separation: The core principle
developed to address image-to-image tasks like
of Support Vector Machines (SVMs) as shown in Figure
segmentation. FCNs replace fully connected layers with
4 can be extended to achieve nonlinear class separation
operations that reverse convolution and downsampling,
using Mercer kernels. This extension allows SVMs to
maintaining full-resolution information. U-Net, a famous
handle more complex data distributions and perform
architecture in medical imaging, employs an encoder-
better when linear separation is insufficient. GANs have
decoder structure for tasks like image segmentation.
been applied to medical image processing. GANs are
used for tasks like data augmentation, generating
synthetic medical images, and transforming images to
enhance certain features. These applications contribute to
improving the quality and diversity of medical image
datasets [17].

Random forests (RFs)

RF is an ensemble learning method that utilize multiple
independent binary decision trees to create a robust
predictive model. The final output is the sum of the
choices made by each decision tree. Each decision tree
works as a simple model (binary classifier) and comes to Fig. 5. Typical architecture of CNN [20]
its conclusion. The method has built-in ways to vote and In medical imaging, CNNs as shown in Figure 5 are
choose which features to use. The algorithm generates extensively used for supervised tasks such as detection,
numerous simple data representations, applies feature localization, classification, and segmentation. Notably,
selection to identify the most informative features, and CNN-based algorithms have outperformed classical
then combines the decisions of selected classifiers methods in competitions and challenges. For instance,
through majority voting. RFs were initially applied to CNN-based algorithms have achieved remarkable
tasks like organ localization and delineation. Over time, accuracy in tasks like liver ultrasound tracking and head
their applications expanded to include tasks such as and neck auto-segmentation, surpassing previous

Authorized licensed use limited to: VIT University. Downloaded on May 13,2025 at 10:21:32 UTC from IEEE Xplore. Restrictions apply.
approaches. These successes have solidified CNNs as the able to connect the two domains.
backbone of cutting-edge algorithms.
VI. CHALLENGES AND
OPPORTUNITIES

Medical imaging datasets can be limited in size and

quality, making it challenging to train robust models.
Deep learning models as shown in Table 1 can be
complex "black boxes," making interpreting their
conclusions harder. Black boxes are crucial in medical
contexts. Models trained on one type of medical imaging
might not generalize well to other modalities or patient
populations. Integrating AI into medical practices
Fig. 6. Structure of Generative Adversarial Networks (GANs) [25] requires regulatory approval, which can be a lengthy and
complex process; ensuring patient privacy and data
Generative adversarial networks (GANs) security and addressing biases in data and algorithms are
Generative Adversarial Networks (GANs) as shown in essential.
Figure 6 revolve around the interaction of two networks: Opportunities: ML can help in the early detection and
a generator (denoted as ℊ) and a discriminator. The diagnosis of diseases by analyzing subtle patterns in
generator (ℊ) aims to map a random input distribution to medical images. ML can aid in tailoring treatment plans
a predefined data distribution through an iterative process, for individual patients based on their unique medical
effectively generating new data instances. In parallel, the imaging data.
discriminator (D) assesses the generated data, and, based
on its assessment, the generator seeks to minimize the TABLE I. STUDY OF SOME EXISTING TECHNIQUES
dissimilarity between the two distributions. This process Author Methodology Advantages Disadvantages

leads to the creation of data that closely resembles the

Litjens (CNN) High Require large
input distribution. The ultimate goal is for the generator et al. accuracy in amounts of labelled
to deceive the discriminator into wrongly classifying the 2017[20] image data.
generated data as authentic. Both networks are trained classification
simultaneously, creating a competitive dynamic where tasks.
Esteva Transfer Quick Performance is
the generator (ℊ) aims to outsmart the discriminator (D)
et al. Learning adaptation to heavily dependent
by generating increasingly convincing data. In contrast, 2017[21] with CNNs new imaging on pre-trained
the discriminator (D) hones its ability to differentiate tasks. model quality.
between real input and generated data. The discriminator Shen et Generative Can generate GAN training can
and generator are typically implemented as Convolutional al. Adversarial synthetic be unstable and
Neural Networks (CNNs) and trained adversarially. 2018[22] Networks medical challenging.
In contrast to Convolutional Neural Networks (CNNs), (GANs) for images for
image data
which have a historical foundation dating back to the synthesis augmentation.
1980s, adversarial learning is a relatively modern
concept. Nonetheless, it has gained rapid prominence
within medical imaging, leading to a substantial increase Yan et Recurrent Effective in Prone to
al. Neural processing vanishing/exploding
in publications in recent times [19]. Initial Generative 2018[23] Networks sequences, gradient problems.
Adversarial Network (GAN) architectures faced (RNNs) for e.g., time-
challenges like unstable training. However, the field of sequential series data.
computer vision has spurred significant advancements by medical data
modifying the architectures of the discriminator (D) and
generator (ℊ) networks or exploring novel loss functions.
The augmentation of control over data generation in Irvin et Dense Net- Can handle May require fine-
GANs involves introducing additional information about al. based models varying image tuning for different
2019[24] for Chest sizes due to imaging modalities.
desired output characteristics, such as real image Radiograph its fully
examples or labels. It has given rise to conditional GANs Interpretation convolutional
(c GANs), supervised learning that relies on matched nature.
training pairs. However, from our point of view, one of
the most essential things about GANs is that they can
learn from semi-supervised or completely uncontrolled VII. CONCLUSION
data sets. GANs are becoming more essential tools,
especially in medical imaging, where well-aligned and The article offers an overview of machine learning,
labelled picture pairs are often insufficient to work. In this specifically its application in medical imaging analysis.
situation, cycle GANs [19] have become more critical The article focuses on fundamental methodological
because they can map two areas in both directions using concepts and demonstrates how advanced machine
unpaired input data. With this function, they have been learning (ML) techniques can contribute to augmenting

Authorized licensed use limited to: VIT University. Downloaded on May 13,2025 at 10:21:32 UTC from IEEE Xplore. Restrictions apply.
various phases of clinical practice. This knowledge can neural networks. Nature 2017;542:115–8.
13. Lotter W, Diab AR, Haslam B, Kim JG, Grisot G, Wu E, et al.
help demystify AI, inspire innovative research directions,
Robust breast cancerdetection in mammography and digital breast
and facilitate its incorporation into clinical settings if it is tomosynthesis using an annotation-efficient deep learning
incorporated into the medical field and made available to approach. Nat Med 2021;27:244–9.
healthcare professionals. Much of the initial research 14. Nikolov S, Blackwell S, Zverovitch A, Mendes R, Livne M, De
Fauw J, et al. Deep learning to achieve clinically applicable
sought to translate recent ML and DL advancements in
segmentation of head and neck anatomy for radiotherapy. arXiv
computer vision to medical contexts, demonstrating the [csCV] 2018.
viability and potential of these methods for enhancing 15. Shen L, Margolies LR, Rothstein JH, Fluder E, McBride R, Sieh
clinical practice. Several papers show how to use CNNs W. Deep learningto improve breast cancer detection on screening
mammography. Sci Rep 2019;9:12495.
for organ segmentation and dose prediction in
16. Komura D, Ishikawa S. Machine learning methods for
radiotherapy or GANs for modality conversion, which are histopathological image analysis. Comput Struct Biotechnol J
examples of this technical transfer. Several other studies 2018;16:34–42. https://doi.org/10.1016/j.csbj.2018.01.001.
have mentioned these books and articles. The next wave 17. Perone CS, Ballester P, Barros RC, Cohen-Adad J. Unsupervised
domain adaptation for medical imaging segmentation with self-
of AI methods designed for medical applications will
ensembling. NeuroImage 2019;194:
emerge once the medical community actively accepts AI 18. Shao S, Wang P, Yan R. Generative adversarial networks for data
technology and adds domain-specific knowledge to the augmentation in machine fault diagnosis. Comput Ind
most advanced AI methods. It will only happen if the 2019;106:85–93. https://doi.org/10.1016/j.compind.2019.01.001.
19. Sandfort V, Yan K, Pickhardt PJ, Summers RM. Data
medical community widely uses AI technology. This
augmentation usinggenerative adversarial networks (CycleGAN)
integration could entail adding domain-specific to improve generalizability in CTsegmentation tasks. Sci Rep
information to model inputs or using specialized loss 2019;9:16884.
functions during training. Several initiatives have already 20. Litjens, G., Kooi, T., Bejnordi, B. E., Setio, A. A. A., Ciompi, F.,
Ghafoorian, M., ... & van der Laak, J. A. (2017). A survey on deep
investigated these possibilities.
learning in medical image analysis. Medical Image Analysis, 42,
In conclusion, AI techniques, specifically ML, have made 60-88.
significant strides, showing how they can help doctors do 21. Esteva, A., Kuprel, B., Novoa, R. A., Ko, J., Swetter, S. M., Blau,
their jobs better. Nevertheless, for healthcare workflow H. M., & Thrun, S. (2017). Dermatologist-level classification of
skin cancer with deep neural networks. Nature, 542(7639), 115-
integration to work well, people from different fields must
118.
work together. This joint effort will set the stage for the 22. Shen, D., Wu, G., & Suk, H. I. (2017). Deep learning in medical
next generation of robust, easy-to-understand, and image analysis. Annual Review of Biomedical Engineering, 19,
powerful AI methods that could change the healthcare 221-248.
23. Yan, K., Wang, X., Lu, L., & Summers, R. M. (2018).
business.
DeepLesion: Automated detection and segmentation of lung
cancer metastases in CT images. In Proceedings of the IEEE
REFERENCES Conference on Computer Vision and Pattern Recognition (CVPR)
1. Cui S, Tseng H, Pakela J, Ten Haken RK, El Naqa I. Introduction (pp. 9290-9299).
to machine and deep learning for medical physicists. Med Phys 24. Irvin, J., Rajpurkar, P., Ko, M., Yu, Y., Ciurea-Ilcus, S., Chute, C.,
2020. https://doi.org/10.1002/mp.14140. ... & Ng, A. (2019). Chexpert: A large chest radiograph dataset
2. Holman JG, Cookson MJ. EXpert systems for medical with uncertainty labels and expert comparison. In Proceedings of
applications. J Med EngTechnol 1987;11:151–9. the AAAI Conference on Artificial Intelligence (AAAI) (Vol. 33,
3. Haug PJ. Uses of diagnostic expert systems in clinical care. Proc No. 01, pp. 590-597).
Annu SympComput Appl Med Care 1993:379–83. 25. Qi An 1 , Saifur Rahman 1 , Jingwen Zhou 1 and James Jin Kang
4. Miller RA. Medical diagnostic decision support systems–past, 2,* A Comprehensive Review on Machine Learning in Healthcare
present, and future:a threaded bibliography and brief commentary. Industry: Classification, Restrictions, Opportunities and
J Am Med Inform Assoc 1994;1:8–27. Challenges Sensors 2023, 23, 4178.
5. Buchanan BB, Buchanan BG, Buchanan BG, Shortliffe EH, https://doi.org/10.3390/s23094178
Heuristic S. Rule-based expert systems: the MYCIN experiments
of the stanford heuristic programming project. Addison Wesley
Publishing Company; 1984.
6. Aikins JS, Kunz JC, Shortliffe EH, Fallat RJ. PUFF: an expert
system forinterpretation of pulmonary function data. Comput
Biomed Res 1983;16:199–208.
7. Miller RA, Pople Jr HE, Myers JD. Internist-1, an experimental
computer-baseddiagnostic consultant for general internal
medicine. N Engl J Med 1982;307:468–76.
8. Buchanan BG. Can Machine Learning Offer Anything to EXpert
Systems? In: Marcus S, editor. Knowledge Acquisition: Selected
Research and Commentary: A Special Issue of Machine Learning
on Knowledge Acquisition, Boston, MA:Springer US; 1990, p. 5–
8.
9. Su MC. Use of neural networks as medical diagnosis expert
systems. Comput BiolMed1994;24:419–29.
10. Aizenberg IN, Aizenberg NN, Vandewalle J. Multi-valued and
universal binary neurons 2000. https://doi.org/10.1007/978-1-
4757-3115-6.
11. Liu X, Faes L, Kale AU, Wagner SK, Fu DJ, Bruynseels A, et al.
A comparison ofdeep learning performance against health-care
professionals in detecting diseasesfrom medical imaging: a
systematic review and meta-analysis. Lancet
DigitHealth2019;1:e271–97.
12. Esteva A, Kuprel B, Novoa RA, Ko J, Swetter SM, Blau HM, et
al. Dermatologist-level classification of skin cancer with deep

Authorized licensed use limited to: VIT University. Downloaded on May 13,2025 at 10:21:32 UTC from IEEE Xplore. Restrictions apply.