Artificial intelligence in medicine: Current applications in cardiology, oncology, and radiology

Abstract

In this article, artificial intelligence (AI) usage and its benefits in medicine are reviewed in the oncology, radiology, and cardiology fields. The relevant literature was searched in PubMed and Google Scholar using the words “Artificial Intelligence”, “Artificial Intelligence in Medicine”, “Artificial Intelligence in Cardiology”, “Artificial Intelligence in Oncology”, and “Artificial Intelligence in Radiology” for the last 10 years. This article covers the AI’s current implications in daily practice, discussing its advantages and disadvantages based on the findings. AI’s effect in medicine for reducing workload, diagnosis, time management, and drug dosing is going to be reviewed especially in radiology, oncology, and cardiology fields as well as general usage of AI in addition to important highlights. Lastly, this minireview evaluates the current challenges of AI technology in medicine and how clinicians should work with this emerging technology.

Key Words: Artificial intelligence; Medicine; Radiology; Oncology; Cardiology; Management; Diagnosis

Core Tip: Artificial intelligence (AI) is the ability of a digital computer system to perform tasks commonly associated with human beings. In this article, AI usage and its benefits in medicine and its specialties are reviewed.

INTRODUCTION

Artificial intelligence (AI) is the ability of a digital computer system to perform tasks commonly associated with human beings. This term is often applied to systems that can reason, discover meaning, generalize, or learn from experience like humans[1]. In 1950, Alan Turing developed the Turing Test, thus making the foundation of AI technology and is being remembered as the “father of AI”[2]. Since the Turing Test, there has been massive development in AI with emerging technologies and methods such as machine learning (ML), convolutional neural networks (CNN), deep learning (DL), and large language models. Especially with the introduction of the ChatGPT in November 2022 by OpenAI, AI technology has made its way into our everyday life from asking simple questions to trying to manage complicated situations. ChatGPT has become the fastest growing internet app gaining 100 million users in two months[3].

In medicine, AI was first described as the MYCIN program in 1977 by Stanford University, using a rule-based expert system to help with therapy selection and dose calculation in cases of bacteremia[4]. Since MYCIN, AI algorithms in medicine have evolved to complex and detailed structures for clinical use. Between 2013 and 2016, AI was most used in diagnostic imaging, genetics, electrodiagnosis, and monitoring[5].

AI has diverse applications differing from operating rooms to emergency procedures or appointment-making to diagnosing patients, all in favor of a much elevated level of efficiency in daily practice. AI can do routine procedures such as data processing, dosage calculation or multi-drug combinations, and immunization schedules for children so that doctors’ workload can be lighter and they can work more on creative things to make developments in their jobs and communicate with patients[6]. Since experience is a very important factor for doctors to make the right decisions when treating a patient or for surgeons to perform surgeries more accurately, AI can work on various data and can help reduce the factor of experience when it comes to doctors’ performances[7]. When examining a patient, AI can work on patients’ data to search for possible diagnoses by making entries of patients’ complaints and assisting doctors[8] (Figure 1). Especially in emergency departments, acting fast and seeing patients in the most efficient way is very important. With AI, the average length of stay can be dramatically reduced. Also, sources and staff can be used more efficiently[9].

Figure 1 Benefits of artificial intelligence in medicine. When examining a patient, artificial intelligence can work on the patient’s data to search for possible diagnoses by making entries of the patient’s complaints and assisting doctors.

Another issue that affects doctors’ work efficiency is malpractice cases. Fear of these cases can make doctors work more cautiously and this can affect the quality of the service that they are going to give and jeopardize patients’ lives. With the help of AI, malpractices can be foreseen and therefore prevented. AI can calculate the possibility of malpractice and show alternative ways of treatment to avoid it. Also, AI can present non-biased evidence when it is needed in malpractice cases[8].

AI can be a tool in a very broad spectrum, reducing morbidity such as diagnosing diabetic retinopathy[10] or helping with higher healthcare costs and mortality by defining nosocomial infections[11], which can affect 100 million patients per year[12]. Diabetic retinopathy is affecting 34.6% of the diabetic patients and is a huge reason for morbidity. AI has demonstrated success in early diagnosis of diabetic retinopathy. Researchers have programmed a DL algorithm for scanning retinal fundus images for diabetic retinopathy and the operating point selected for high sensitivity. Optimized for high sensitivity, the algorithm achieved 97.5% and 96.1% sensitivity and 93.4% and 93.9% specificity across the two validation sets[6]. Nosocomial infections are a reason for mortality and morbidity in hospitals. Also, they raise the cost of health care rapidly. All around the world, nosocomial infections affect an estimated 100 million patients per year[7]. Therefore, with being a serious and global problem, diagnosing and recognizing nosocomial infections early is very important. AI can help monitor patients’ vital signs and warn health workers about nosocomial infections. In a study, researchers combined medical knowledge with fuzzy theory AI method and created a monitoring and infection identification system called Moni/Surveillance in Vienna General Hospital. The usage of this AI system showed promising results[8].

In this article, we will review the applications of AI as well as its limitations which are derived from these technologies, mainly focusing on the areas of cardiology, oncology, radiology, allergy-immunology, and infectious diseases. The relevant literature was searched in PubMed and Google Scholar using the key words “Artificial Intelligence”, “Artificial Intelligence in Medicine”, “Artificial Intelligence in Cardiology”, “Artificial Intelligence in Oncology”, and “Artificial Intelligence in Radiology” for the last 10 years (published between April 2015-April 2025). Articles with titles that do not include these keywords are not screened.

AI IN CARDIOLOGY

Cardiologic diseases are mostly mortal, morbid, chronic, heterogeneous, slow, silent, and regressive but with rapid and life-threatening consequences. Therefore, AI methods such as DL, CNN, linear regression, or tree-based decisions have much potential in cardiology[13].

In cardiac evaluation, echocardiography and electrocardiography are essential but interpretation of the received data from these methods can be difficult because of the artifacts and errors. Especially, detecting conditions such as premature beats, atrial fibrillation, ventricular hypertrophy, non-sinus rhythms, and pacemaker dysfunction is very difficult. Thus, clinicians can use the help of technologies derived from AI. With this purpose, researchers from the Mayo Clinic have developed an AI algorithm using a CNN called AI-ECG. The study included 720978 adult patients aged 18 years or older with a standard 12-lead electrocardiography (ECG) performed at the Mayo Clinic ECG laboratory between 1993 and 2017. These ECG samples have been re-evaluated with AI-ECG for determination of primary and secondary rhythms, axis deviation, and chamber enlargement, and detection of atrioventricular and intraventricular conduction delay, myocardial ischemia, and waveform abnormalities. AI-ECG showed promising results in researched conditions[14]. Also, another study using a Food and Drug Administration (FDA)-approved AI system called Viz.ai showcased promising results in diagnosing hypertrophic cardiomyopathy (HCM). HCM is one of the main causes of sudden cardiac death in youth and its prevalence is approximately 1 in 500. Early diagnosis is the key to preventing death. Viz.ai has identified HCM with a sensitivity of 68.4% and specificity of 99.1%[15]. Therefore, AI has much potential to reduce artifacts and help clinicians make more accurate comments and diagnoses based on ECG data.

Ultrasonographic echocardiography is also commonly used in cardiology but there is a high chance for error because the quality of the echocardiogram image is influenced by the clinician’s experience and competence. It is harder to present generalized data in echocardiography. Therefore, AI models have been developed regarding this issue. AI can help reduce noise and improve navigation of the ultrasonography probe. Some of AI models have already been approved by the FDA in the United States of America[16].

Cardiac magnetic resonance imaging (MRI) with late gadolinium enhancement (CMRIG) is a technique used for identifying ischemic or fibrotic tissue in the heart. It can be used in the diagnosis or management of myocardial infarction, myocarditis, sarcoidosis, amyloidosis, hypertrophic cardiomyopathy, and diabetic cardiomyopathy. The biggest disadvantage of this technique is the need for intravenous (IV) contrast agent injection. Contrast agents cannot be used in patients with kidney failure or pregnancy due to the increased risk of nephrotoxicity. Also, gadolinium administration raises the cost of the procedure significantly. To overcome these obstacles, a virtual native enhancement (VNE) imaging technology was developed using AI. In a study conducted on hypertrophic cardiomyopathy patients, VNE showed better quality images than CMRIG without the need for an IV contrast agent with less cost. VNE has much potential to replace CMRIG and to help reduce the side effects caused by contrast agents[17].

Different from other fields, cardiology uses biomarkers such as cardiac troponins, creatine kinase, D-dimer, myoglobin, and cardiac enzymes. These biomarkers are essential for diagnosis, treatment, and acute management. AI can help use biomarkers more effectively when assessing each patient individually[18] and choosing the right biomarker for the highest specificity and sensitivity while classifying the data according to clinical cases[19].

Also, cardiac conditions and treatments need to be evaluated for a long time. Patients must come in regularly and the clinicians must modify the treatment by calculating the patients’ overall health. AI can help clinicians when planning the right and individualized treatments[20] and assessing the patient’s condition when deciding on an invasive procedure[21].

AI IN ONCOLOGY

Cancer is a worldwide and serious health problem. Cancer incidence is rising with the elevated levels of radiation in daily life, pollution, and chemicals. In 2022, cancer prevalence in Asia is 6.5 million and in Europe is 3.5 million. And by the time of 2050, its prevalence was predicted to almost double[22] (Figure 2).

Figure 2 Estimated numbers of cancer cases in 2050. By the time of 2050, cancer prevalence was predicted to almost double in Asia and in Europe.

With this alarming number of cases, early identification of cancer is more important. Considering that most of the cancer cases are asymptomatic, clinicians can use the help of technologies like AI. Also, when patients are diagnosed, they need excessive care and treatment over a prolonged period of time. AI can help with the diagnosis, choosing the right treatment regime, and following the patient’s health data.

For the diagnosis of cancer, imaging is a must. Especially, ML and image processing methods have much promise in this area due to examining mass amounts of screenings for abnormal findings and detecting them more accurately with cross-referencing. AI can be helpful in cancer imaging with monitoring, detection, segmentation, classification, and characterization of the lesions. Different from MRI and computed tomography (CT), AI can offer more qualitative data rather than quantitative data[23]. In cancer imaging, AI has a great prospect in a very broad spectrum from helping clinicians identify early gastric cancer in endoscopy images[24] to generating three-dimensional (3D) images from positron emission tomography (PET) screening[25], all for better understanding of the cancer diagnosis and identifying malignant masses as soon as possible to reduce the mortality. These new technologies can be the key to understanding cancer’s pathophysiology and carrying the treatments into the next step[23] (Figure 3).

Figure 3 Artificial intelligence in cancer screening. New artificial intelligence technologies can be the key to understanding cancer’s pathophysiology and carrying the treatments into the next step.

Lung cancer is one of the deadliest cancers of all time and the number of new cases estimated is very high, along with breast cancer[26]. Thus, there has been great interest in using new tools like AI for the identification and classification of lung cancer.

The first step of lung cancer diagnosis is scanning chest radiograph images and detection of nodules. After detection, clinicians can suspect lung cancer. AI-assisted image analysis with ML has demonstrated efficacy in detecting nodules that might otherwise escape radiologists' detection[27]. After detection, more accurate classification of the nodules or assessing the genetic figuration of the mass can carry therapies like immunotherapy into the next level. Usually, a biopsy is used to differentiate tumor type, but tumors can be heterogeneous. Therefore, biopsy is not always reliable considering taking sample tissue from only one area[23]. More overall knowledge about the assessment of the mass can give clinicians the needed insight into the patient’s status. In a study, a DL algorithm has been trained to extend the assessment of nodules rather than using the traditional classification such as solid, part-solid, non-solid, and calcified nodules[28]. AI has brought depth when looking into the characters of nodules, thus making the way to improving the selected therapy’s outcome.

Also, immunotherapy is often used in lung cancer, and it is chosen based on the tumor’s gene expression. ML algorithms can be trained using pathology reports and can define different genetic expressions in lung cancer easily. It is a huge step for making better-personalized immunotherapy. It can lighten clinicians’ diagnostic burden while improving the prediction of cancer susceptibility, mortality, and recurrence up to 25%[29].

Breast cancer is the most common cancer in the United States and it has the highest estimated chance of incidence[26]. Mammography is a useful tool for screening breast cancer. It can detect clusters, masses, or calcifications. And with BIRADS classification, clinicians can have an insight into the prognosis. Although mammogram findings must be proven with other screening methods, it is still the key to diagnosing breast cancer[30].

In a study, CNN has been used to classify the breast tissue as normal or mass tissue. CNN has been trained using 168 biopsy-proved breast cancer and 564 normal breast images approved by experienced radiologists. The study showed that CNN worked with a true positive fraction of 90% and false positive fraction of 31%[31]. Besides that, there are multiple software programs showing promising results in identifying pathological findings and diagnosing breast cancer[32]. Also, when evaluating a mammogram, increased density of breast parenchyma is an issue. Because increased density can hide masses and make it more difficult for the clinician to interpret the mammogram more accurately. Breast density increases as people age, which is another risk factor for breast cancer. DL can grasp a better understanding of identifying masses in more dense parenchyma and shows much potential for improving the diagnosis of patients[33].

Mammography is also used in screening. Screening programs for breast cancer identification exist worldwide and many clinicians perform these programs[34]. Screening programs with mammograms provide the chance for earlier diagnosis and early start for adjuvant therapy, hence reducing the mortality rate up to 68%[35]. However, these often usage of mammograms can have negative outcomes. In a study, researchers showed that screening with mammography leads to 132 over-diagnosed breast cancer cases in 10000 women[36]. Excessive usage can also lead to misusage of sources and increased workload for clinicians. Therefore, with these situations, an opportunity for AI arises for reducing the workload and over-diagnosis. Triaging the mammography findings with AI software and consulting the radiologist on only needed cases is reducing the workload and makes the performance of the clinicians better. AI can grade the findings as more or less likely to be diagnosed as cancer and can work coherently with the clinicians[37,38]. However, implementing AI does not always bring positive outcomes. In a study, evaluating mammograms with computer-aided detection is associated with an increased rate of biopsy but on the contrary, reduced accuracy of findings[39]. AI shows great potential in interpreting mammograms, but it should be used as a tool, work coherently to enhance the performance of clinicians, and not be overthrown or ignored when it comes to making decisions about patients’ care.

In oncology, AI is mainly being used in cancer diagnosis, but it is surely not the only area for it. AI can detect metastasis, evaluate the genetic or hormonal structure of the mass, or predict the patient’s survival and recovery.

In a study, a neural network model was accomplished to determine the isocitrate dehydrogenase (IDH) mutation status of gliomas from MRI with an accuracy of 82.8%. IDH mutation means a longer chance of survival in glioma patients and can potentially alter the treatment of choice. With neural networks, IDH status can be assessed without invasive procedures[40]. Another prognostic factor for gliomas is the deletion of chromosome arms of 1p/19q. The existence of this deletion is linked with better therapy response and a higher chance of survival without relapses. Best performing CNN predicted the chromosome deletion with a 93.3% sensitivity and 82.22% specificity from MRI[41]. With this accomplishment, deletion status can be assessed without biopsy, creating a chance for more accurate treatment and increasing the life expectancy in low-grade glioma patients.

Another factor for the prognosis of cancer patients is whether the primary lesion metastasizes or not. Especially in breast cancer due to anatomical relatedness, axillary lymph node metastasis is a very important prognostic factor and helpful for assessing the mortality or predicting the chance of metastasis in other regions[42]. In a cross-sectional study that evaluated 32 DL algorithms, 7 algorithms showed greater discrimination of metastatic tissue in biopsy samples than a panel of 11 pathologists[43].

In addition to these points, artificial neural network algorithms have proven effective in predicting delayed discharge or readmission after laparoscopic colorectal surgery. With these results, neural networks have the potential to create an ideal predictive tool for postoperative patient recovery[44].

AI IN RADIOLOGY

Since the discovery of the X-ray technology by Roentgen, imaging technologies have been a huge part of medicine. With the advancement in technology, various imaging techniques such as PET, CT, and MRI are being used by clinicians around the world. Thus, radiology is one of the fundamental specialties in health care. Clinicians use imaging for diagnosis, screening, examining the patient, or selecting the therapy. In 2023, 45 million imaging tests were reported in the United Kingdom, with an increase rate of 2.2%[45]. With the increasing number of imaging units and the ordered scans[46], radiologists face a variety of problems. They experience higher rates of burnout, unreasonable amount of workload, and inadequate workflow efficiency[47]. On top of that, there have been several reports from various countries in the past years about radiologist shortage in medical care[48-50].

With the launch of ImageNet in 2009, AI technologies have reached a new level when it comes to classifying, identifying, and segmenting visual databases, as well as recognizing patterns and associating them as certain indicators. ImageNet can also function as a pool for annotated datasets for CNN training[51]. Therefore, cooperation with AI became inevitable for helping radiologists. Between 2015 and 2019, over ten thousand papers have been published regarding AI and radiology with cancer imaging being the most focused area[52]. There are currently 48 AI programs for abdominal imaging, 46 for breast imaging, 67 for cardiac imaging, 91 for chest imaging, 41 for musculoskeletal system imaging, and 112 for neuroradiology, which were all approved by the FDA[52,53] (Figure 4).

Figure 4 Usage of artificial intelligence programs approved by the Food and Drug Administration in radiology. AI: Artificial intelligence.

For the management of the patient in emergency departments, radiological imaging is a must. Clinicians need to act fast to diagnose and triage the patient since it is crucial to patients’ survival. It is important to identify pathologies from images or to eliminate the artifacts by recognizing them and choosing the right procedure of imaging for the right patient. With the need to quickly act the pressure of dealing with mortal situations, and long and overnight working hours, clinicians working in emergency departments are more prone to making wrong or delayed diagnoses. Clinicians seem to make diagnostic errors mostly in patients with fractures (44%), myocardial infarctions (7%), or intracranial bleeding (6%), with 15% of the errors resulting in severe harm or death[54]. Implementing AI in daily practice can help clinicians in most common cases presenting to the emergency departments such as pulmonary embolism, stroke, acute abdominal pain, and fractures[55].

AI tools are open for integration into the diagnosis-making by alerting the clinicians 24/7 when suspicious findings are detected for stroke, intracranial hemorrhage, or pulmonary emboli, hence guiding the clinician for better management of the patient[56]. Chest radiography is one of the most used radiological imaging modalities in emergency departments. Evaluation of the chest radiography most often requires the attention of radiology residents; however, it is not possible to reach that due to working hours or staff shortage. In a study evaluating AI’s accuracy for chest radiography, AI performed equally well with radiology residents but outperformed non-radiology residents[57]. AI can work 24/7 in the background reducing the need for radiology residents and lightening the workload. AI technology has proven successful when assisting clinicians in evaluating CT pulmonary angiography by catching 19 pulmonary emboli cases missed by clinicians, therefore acting as a double-checking standard in clinical practice[58].

For musculoskeletal imaging, AI has proven effective for classifying distal radius fractures from frontal and lateral radiographs when it comes to identifying fragment displacement or joint involvement as well as eliminating artifacts that usually stem from bone superimposition or osteoarthritis[59]. Also, using standard measurement techniques for anatomical landmarks such as hip-knee angle, tibia/femur or leg length, and pelvic obliquity is essential when diagnosing musculoskeletal diseases. Automating these measurements with AI is a huge step forward saving workforce and energy[60].

The reduced quality of radiographic imaging can be a huge problem for clinicians. DL programs have been trained to classify MRI’ quality as acceptable or unacceptable[61]. For MRI, it is even possible to assess fetal MRI which artifacts are most likely to occur due to limitations[62], giving insight to the clinician on how to evaluate or even decide to dismiss the image.

PET is a widely used imaging technique with proven clinical benefits. PET imaging has extended its benefits with the synergy between new technologies such as the introduction of PET/CT scanners. PET image reconstruction is another method for PET images to be open for interpretation and ready for clinicians. With PET image reconstruction, cross-sectional images that stem from radiotracer distribution can be provided as well as resolution of the images can be enhanced. There are several mathematical formulations for this cause[63]. The potential in DL and CNN for doing monotone mathematical calculations in rapid speed and accuracy while being open for training and modification, made AI a suitable technology for PET image reconstruction in the path of getting a more accurate radiological image. For this purpose, different AI algorithms have been introduced with the technology of CNN or DL. These algorithms differ from each other in the aspects of interpretability, suitability for 3D image reconstruction, inference time, generalization, and expected image quality if within training distribution. Each algorithm has its advantages and disadvantages[64]. Especially, DL algorithms have been found successful when it comes to metal artifact reduction that can stem from prosthetics or implants. Different DL and metal artifact reduction algorithms can be combined for better image reconstruction for PET images or to enhance the quality of CT imaging[65,66]. However, the need for different algorithms for different imaging techniques such as CT or MRI, or for different sole purposes such as artifact reduction raises the problem of instability. AI algorithms can produce other artifacts stemming from programming or training data sets, making it difficult to reach a generalized universal image quality[67]. CNN-based PET image reconstruction programs may produce more artifacts. On the other hand, DL-based systems are harder to program with needing larger training data sets. Therefore, the need for new technologies and emerging AI algorithms is still there. A new algorithm integrating a deep image prior framework has been introduced, and it is superior to DL reconstruction programs and eliminates potential biases that arise from different training data sets[68].

AI in radiology is not only effective in artifact reduction or image reconstruction but also in the synthesis of images. With the introduction of cycle-consistent adversarial networks[69] and U-Net[70] algorithms, AI for image synthesis has reached a new level. There are several DL algorithms for the synthesis of pseudo-CT, synthetic MRI, and synthetic PET, and they all show great promise in medical usage[71]. Even though they have their advantage and disadvantages, a study conducted comparing cycle-consistent adversarial networks and U-Net algorithms in pseudo-CT/MRI synthesis showed that the U-Net algorithm showed superior performance, especially in pseudo-CT[72]. These CNN and DL algorithms were found to be successful for image synthesis and also applicable for more vulnerable patient groups such as pediatric patients. A study comparing CT and synthetic CT (sCT) quality in bone elements of sacroiliac joints in children found that there is no significant lowering in quality between sCT and CT. On top of that, sCT performed even better when it comes to showing cortical delineation[73]. DL also showcased the potential to reduce given radiation dose to patients by up to 75% with deionizing algorithms[74]. AI has a key potential for the next step of radiological imaging. They are also adaptable and changeable to need, such as reprogramming the algorithm[75] or adapting to different imaging techniques such as fluoroscopy[76] (Figure 5).

Figure 5 Artificial intelligence use in various imaging technologies. CT: Computed tomography; MRI: Magnetic resonance imaging; PET: Positron emission tomography.

Another usage of imaging technologies such as MRI and CT is to plan radiotherapy. Clinicians evaluate the lesions with CT and MRI and decide the radiotherapy protocol and dosing. However, working with two different imaging techniques comes with errors and interpretation mistakes. To reduce the workload, cost, and radiation doses, clinicians are thinking about a new radiation therapy that uses solely MRI for treatment planning. When compared to procedures including CT scans, MRI-only protocol did not show any difference in the outcome of treatment planning and has a chance of replacing CT protocols[77]. On the other hand, with the development of AI technologies and not wanting to lose the advantages of CT imaging, researchers started using sCT from MRI in radiation oncology. sCT technology is applicable and useful for clinical use, eliminating unnecessary radiation doses as well as making it possible to evaluate the patient’s condition from a broader perceptive. sCT algorithms apply in radiation oncology from brain to prostate lesions[78] or pediatric groups as well as adult patients[79].

AI shows great promise in cross-referencing studies; however, it seems that further study for the correct implementation of AI in clinical practice is much needed. According to an article, radiologists who use AI technologies on daily basis, experience higher burnout than the ones who do not. These rates have been associated with the higher expectancy of work performance from the radiologists who work with AI[76]. In a survey conducted by the European Society of Radiology, 47.6% of the responders do not use AI and 30.4% do not use AI at present but plan on using it in their daily practice. In the same survey regarding their workload status with AI in their practice, 51% of the responders accept increase while 49% accept decrease[77]. When evaluated, the impact of ML-based clinical decision algorithms has a positive impact on performance, especially in less experienced residents’ daily practices compared to their senior colleagues, and overall AI does not make significant changes in 80% of the cases[78]. These data can derive from the fact that AI technologies need adaption for medical use. Implementing data processing technologies to medical images remains challenging. With the differences between nature images and medical images, scientists must find new ways to program CNN for optimal findings. Training CNN through transfer learning from large scale data has proven effective when researched on identifying lymph nodes and interstitial lung disease. With that, training the CNN from scratch or exploring different properties of training also has potential for achieving better outcomes[79].

AI APPLICATIONS IN OTHER MEDICAL FIELDS

Even though we focus mainly on the applications in cardiology, oncology, and radiology fields (Table 1), AI has much broader clinical applications in medicine. In this chapter, we will briefly highlight its applications in allergy- immunology and infectious diseases.

Table 1 Purposes of using artificial intelligence in fields of cardiology, oncology, and radiology.

Cardiology	Pathology	Radiology
Detecting rhythm and wave abnormalities in the electrocardiography	Lesion segmentation and classification	Reducing artifacts and noise
Diagnosing hypertrophic cardiomyopathy	Detection of CT nodules in the lungs	Synthesizing pseudo-images
Noise reduction in echocardiography	Assessing hormonal and genetic status of lesions	PET image reconstruction
Cardiac MRI using VNE technology	Improving mammography sensitivity	MRI radiotherapy programs only
Cardiac enzyme classification	Prognostic prediction and staging	Emergency room clinical assistance
Cardiac resynchronization therapy adjustment	Detection of metastases	Standardization of anatomical landmark measurements

MRI: Magnetic resonance imaging; VNE: Virtual native enhancement; CT: Computed tomography; PET: Positron emission tomography.

AI IN ALLERGY-IMMUNOLOGY

Allergy-immunology is a very sophisticated field with a large set of diseases, dealing mostly with chronic and multi-systemic conditions, thus attending both pediatric and adult patients. Allergic diseases are considered an umbrella term containing several diseases (rhinitis, urticaria, asthma, atopic dermatitis, etc.), and differentiating one from another is not always easy in clinical practice. Allergic and immunologic diseases have similar pathogenic pathways including cytokines like IL-5, IL4R, JAK1, and NR3C1. Identifying these cytokines and their role in pathogenesis with AI-driven targeting systems can open a new chapter in targeted immunotherapy and more accurate diagnosis[80]. AI research in allergy-immunology mainly focused on asthma, inborn errors of immunity (IEI), and atopic dermatitis. However, research on other conditions such as anaphylaxis, urticaria, and eosinophilic gastrointestinal disorders is emerging[81]. In asthma diagnosis, the main challenge is the differentiation from chronic obstructive pulmonary disease (COPD). Pulmonary function tests are not specifically discriminative, especially when it comes to diseases like asthma-COPD overlap syndrome. Recurrent neural networks have proven to differentiate asthma from COPD with 68% accuracy and can work when complexity such as asthma-COPD overlap syndrome is presented[82]. In the prognosis of asthma, the main cause of morbidity and mortality is asthma exacerbations. Exacerbations need immediate care and diagnosis. The main solution for reducing exacerbations is the right prophylactic therapy and drug dosing by predicting the prognosis and quick evaluation. Researchers created a two-rule-based algorithm using electronic health data from patients, and the algorithm proved to identify asthma-related emergency department encounters with 95% specificity and sensitivity[83]. When calculating the patient’s chance of presenting with asthma exacerbations, the Asthma Guidance and Prediction System is considered effective for predicting asthma attacks and reducing the median time of patient evaluation[84].

IEI are one of the causes of morbidity and mortality in children. Early diagnosis in IEI is much needed to protect the patients from “diagnosis odyssey” and ultimately reducing the performed tests for diagnosis while saving time[85]. Today, clinicians use the Joffrey Models 10 Warning Signs (JMF10WS) for suspecting IEI in a patient. However, JMF10WS criteria are still not enough to prevent delayed or underdiagnosis. Different AI models especially ML models can enhance the sensitivity of screening tests by adding factors such as typical IEI diseases, IgG/IgA/IgM levels, age, body weight, or lymphopenia[86]. Training AI algorithms in broader signs for IEI rather than only sticking to JMF10WS is promising for early IEI diagnosis.

AI IN INFECTIOUS DISEASES

Infectious diseases cause major handicaps in healthcare especially in public health. Prevention of infectious diseases is crucial in society for upholding greater standards of healthcare and prosperity. Clinicians lack new methods to diagnose and treat chronic infections known for years such as human immunodeficiency virus (HIV) infection and tuberculosis (TB) while struggling with antibiotic resistance and infections like severe acute respiratory syndrome coronavirus 2 thanks to modern-age medicine. Hence, there is a need for a new chapter in infectious diseases due to struggles coming from both pharmaceutical and clinical aspects of medical practice. The gold standard for diagnosis in almost every infectious disease is blood culture but incubation time for growth is approximately 24-48 hours. This period forces clinicians to use empiric antibiotics to avoid complications coming from the delayed start of therapy. Empiric antibiotic use increases antibiotic resistance and causes unnecessary drug usage. Predictive DL algorithms have been proven to classify blood samples as contaminated or not contaminated when trained to predict the existence of the most common pathogens (E. coli, Klebsiella, S. aureus, Enterococcus, Streptococcus), making it possible to diagnose bloodstream infections without waiting for a period[87].

TB is still a major health issue and has a unique multiple-phased prognosis, and it is difficult to choose which patient to give prophylaxis or treat. Latent TB (LTB) occurs when a patient does not have active TB (ATB) symptoms but has positive blood tests or tuberculin skin test (TST). The Centers for Disease Control and Prevention recommends annual screening in high-risk groups for LTB[88]. However, conventional skin and blood tests can lead to misdiagnoses because interpretation is hard to standardize. Also, the main reason for screening tests in LTB patients is their elevated chance of developing ATB. Clinicians must be alert in LTB patients so that they can diagnose ATB early and prevent comorbidities, and they need new technologies to detect ATB even before the patient’s symptoms worsen. Diagnostic blood and skin tests have low positive predictive values down to 4.2% (in a blood test) and 3.5% (in a TST) when tested for ATB development percentage in LTB patients[89]. Despite these low predictive values in conventional testing, screening algorithms trained with sputum examination, TB symptoms, and chest radiography reached a 90% sensitivity for ATB screening among LTB patients[90]. Also, with the help of decision-tree-based algorithms, scientists achieved to distinguish peripheral blood mononuclear cells’ gene expression levels based on their sensitivity to differentiate LTB and ATB. The algorithm detected CXCL10, ATP10A, and TLR6 genes to have the highest sensitivity up to 71% and specificity up to 89%[91].

Antimicrobial resistance (AMR) is one of the major challenges in medicine, causing morbidity and mortality. With the excessive use of antibiotics or prophylactic treatments, AMR is rising as a major worldwide health problem. On the contrary, production of novel antibiotic agents is very low, therefore drawing attention to AMR as more of a serious handicap. To reduce AMR or fight with ML algorithms can predict which patient needs pre-exposure therapy for HIV infection[92], being a guideline for which patient to give prophylaxis, thus reducing excessive antimicrobial agent usage. Conventionally, an antibiotic resistance test is performed when a sample is obtained from patient to see if their pathogen is resisting to existing antimicrobial agents. Typically, these processes can take up to a minimum of 24 hours, which can be crucial for choosing a patient accurate treatment. In a study, an ML algorithm can successfully predict antibiotic resistance to the most common pathogens in urinary tract infections when trained based on a patient electronic health data such as demographics, past drug prescriptions, and past urine samples[93], ultimately reducing the time for choosing the most suitable drug prescription in treatment. In novel drug development, DL algorithms predict the genes that cause AMR[94], therefore leading to a more AMR-aimed perspective in the drug development industry. On the other hand, again DL algorithms have proven successful in detecting potential biosynthetic gene clusters[95], which is the main criterion in choosing which nature-derived material can be a candidate for antimicrobial drug development.

There are still so many AI studies in medicine even though we tried to cover as much as possible in certain areas, exploring AI technologies and how they can transform the medical practice. AI algorithms’ applications and literature are broad including every inch of medicine from anesthesiology to nephrology and even psychiatry[96-98].

CURRENT CHALLENGES AND LIMITATIONS OF AI APPLICATIONS

The main challenge of AI in clinical use derives from its training technique. AI technologies like ML and DL need large data sets to function and the selection of these datasets can cause biased or limited outcomes of the program. Still, AI ChatBot’s response derives based on sex or ethnicity, causing misleading or ineffective consultation in therapy selection[99]. Other than biased datasets, AI’s predictive value drops in not common or overlooked diseases because of the lack of existing literature and data. In the training and functioning process, many AI algorithms work in a “black box” protocol, remaining unknown of their inner workings. This raises the issue of transparency and being close to easy modifications or receiving criticism when they make non-coherent or non-cohesive responses. Especially ChatGPT algorithm had a lot of backlashes regarding its tendency to hallucinate, a phenomenon referring to ChatGPT’s unreal or heavily modified responses[100]. These responses can cause harm to the patients, raising the issue of malpractice. When evaluating ChatGPT’s accuracy, 8% of the outcomes have been classified as completely wrong. Even though other outcomes did not cause any harm to the patient, they have the potential to jeopardize doctors’ credibility[101]. Malpractice issues raise the need for regulations and legal limitations of AI use, highlighting governance as the other aspect of the implantation challenge. Other than malpractice, using AI can bring up issues as feud or impersonating. In some studies, AI is considered indistinguishable from human work[102]. It is crucial to bring up generalized regulations to AI systems so that they cannot be used without clinicians’ supervising, therefore preventing misleading medical advice from mistrusted sources.

Implementing data processing technologies to medical images remains challenging. Especially for the use of AI in radiology, image processing algorithms still need broad modifications. With the differences between natural images and medical images, scientists must find new ways to program CNN for optimal findings. Training CNN through transfer learning from large-scale data has proven effective when researched on identifying lymph nodes and interstitial lung disease. With that, training the CNN from scratch or exploring different properties of training also has potential for achieving better outcomes[103]. However, today, the outcome of AI in image processing cannot be fully trusted. In other areas of medicine rather than image processing, this implantation problem remains an issue. Thus, most of the AI experiments or revisions on algorithm modifications are being made in non-clinical environments[104]. Clinicians only reach the final product and have little say during the creation process.

Other than limitations coming from algorithms’ working protocol, implications of AI raise ethical, security, or trust issues. For using AI in clinical practice, patients’ data must be shared with these algorithms. This raises cybersecurity issues and can ultimately violate the sacred trust between the clinician and the patient, which is the key to making the right diagnosis and treatment. Patients still trust their doctors rather than AI and are relieved to know that their doctors are the ultimate decision-makers in their treatment process[105]. Upholding this trust is crucial to maintaining high standards of clinical practice. Therefore, clinicians must not rely too much on AI technologies because it is easy and approachable but make decisions according to their own judgment.

POTENTIAL FUTURE DIRECTIONS AND UNRESOLVED ISSUES

When it comes to clinicians’ point of view, challenges are coming from misinformation or prejudice. Many clinicians think that AI is going to steal their jobs or disqualify them. Especially, senior residents do not report any change in their daily practice coming from AI compared to their junior colleagues[106]. They are frightened to get used to AI technologies. Many healthcare workers must be educated to understand AI or how to use it while getting the most benefit. This needs organization and sources for education, making the transition to AI a much harder process. Some clinicians experience higher burnout rates when using AI[107] and 51% of them believe that AI is going to increase their workload rather than decrease it[108]. This attitude comes from patients’ or hospitals’ expectancy of faster and more efficient workflow when invested in AI.

CONCLUSION

AI is an emerging technology making its way into medicine and clinical use rapidly. In this article, we focused on its impact on the cardiology, radiology, and oncology fields. AI technologies like ML, DL, and CNN show great promise in diagnosis, treatment planning, prognosis assessment, and overall helping the clinician give the best care to the patient. However, AI is not a magical tool that can erase all the problems in medicine or cannot be used as the sole decision-maker. Clinicians must adapt to this new technology and learn to incorporate it into their practice. AI must be reevaluated in the light of clinical practice to overcome limitations. All factors such as patients, government, hospitals, and clinicians must be incorporated into the process. There are a massive number of papers in this area, many of them showing promise in cross-referencing studies. However, there is still much to do for corrective usage of AI in best interests of both the patient and the clinician.