The literature search for this review was performed using PubMed, MEDLINE, Cochrane Library, CINAHL, and Google Scholar for relevant studies from inception until June 2023. An AI research assistant (Elicit) was also used to search for relevant papers using the same terminology. In addition, PROSPERO was searched for ongoing similar systematic reviews. The first and senior authors are experienced in systematically reviewing the literature and have published several reviews. In addition, the authors have consulted experts using Editage services to achieve a high level of reliability. Please see Multimedia Appendix 2 for a detailed search strategy.

In order to achieve the objective of this review, the databases were searched using keywords and their Medical Subject Headings (MeSH) terms connected by the Boolean operators “AND,” “OR,” and “*.”

The search used the following terms and their MeSH terms: artificial intelligence, AI, deep learning, machine learning, cardio-oncology, cardiotoxicity, cardiac toxicity, cancer treatment, cancer therapy, “artificial intelligence,” “machine learning,” “AI augmentation,” “deep learning,” OR “AI” AND “Cardio-oncology,” “Cardiotoxicity,” “Cardiovascular toxicity,” OR “Cardiac toxicity” AND “chemotherapy,” “anthracycline,” “cytotoxic regimens,” “immunotherapy,” “Cancer treatment,” OR “therapy-induced” AND “Imaging,” “Echocardiogram,” “Echo*,“ “Cardiac magnetic resonance imaging,” “CMR,” “Multigated acquisition,” “MUGA,” “Cardiac computed tomography,” OR “CCT.”

Original studies reporting AI outcomes in adult patients diagnosed with any type of cancer and undergoing cardiotoxicity assessment were included. Outcomes included incidence of cardiotoxicity, LVEF, risk factors associated with cardiotoxicity, heart failure, myocardial dysfunction, signs of cancer therapy–related cardiovascular toxicity (CTR-CVT), echocardiography, and CMR. Non-English studies, case reports, literature reviews, studies on children, and studies that did not include CTR-CVT were excluded from this review.

The first, second, and third authors (MR, HM, and AR) independently assessed the included articles according to the 42-item Checklist for Artificial Intelligence in Medical Imaging (CLAIM) [28]. CLAIM is modeled after the Standards for Reporting of Diagnostic Accuracy Studies guideline. It addresses the application of AI in medical imaging, including classification, image reconstruction, text analysis, and workflow optimization [28]. Subsequently, the first, second, and third authors cross-checked each other’s articles, and conflicts were resolved through group discussion.

Finally, both HM and AO independently assessed the risk of bias for each study across ROBINS-I’s (risk of bias in nonrandomized studies - of interventions) 7 domains: confounding, selection of participants, classification of exposures, deviation from intended exposure, missing data, measurement of outcomes, and selection of reported results. Each domain was rated as low, moderate, serious, or critical based on the each domain’s algorithm, with the most severe rating across all domains determining the overall assessment for each study. Any disagreements in the assessments were discussed until a consensus was reached, with one reviewer (DJ) ensuring consistent application of judgments. Additionally, we engaged a fifth-year medical student experienced in systematic reviews and various research projects to independently evaluate the risk of bias for all included studies using the same tool.

According to the CLAIM checklist, the first, second, and third authors (MR, HM, and AR) extracted data from the included studies. All discrepancies were resolved after a discussion, with HM acting as an arbitrator. Descriptive information about each study was recorded, including publication details (author, year, and country), sample size, cancer type, imaging technique, AI model, outcomes, and limitations. AO performed analysis, and figures were generated using RStudio (version 2023.06.0; Posit PBC).

A total of 883 articles were identified in the database search, comprising 593 articles from PubMed, 267 from MEDLINE, 2 from CINAHL, and 21 from Elicit. After eliminating duplicate titles and articles in non-English languages, 617 articles remained. Then, the title and abstract of the 617 articles were screened independently by the first and second authors (MR and HM), and 44 remained. The authors reviewed full texts and 7 articles met the inclusion criteria (Figure 1).

The quality assessment used the 42-item CLAIM. The distribution and percentages of different sections and items of CLAIM compliance are depicted in Figures 2 and 3. These sections include title/abstract, introduction, methods, results, discussion, and other information. Each section is categorized into “No” and “Yes” groups, indicating whether it is reported in the selected articles.

Based on the data, the title/abstract section was compliant in 93% of the articles (1 article was compliant with the title but not the abstract, which was considered as half compliant). An introduction section was included in all 7 articles, representing 100% compliance. Methods had 77% compliance, results represented 77% compliance, and there was a discussion in 93% of the articles, while other information was 76% compliant. Items 5 and 6 of the checklist—specific to the study methodology and design—were met as follows: 5 studies were conducted prospectively, while the remaining 2 were conducted retrospectively. Moreover, 4 studies were reported as feasibility studies, 2 were exploratory studies, and 1 was a model creation study. Finally, items 10 and 27 of the CLAIM criteria did not apply to the 7 studies.

In total, 7 studies conducted between 2018 and 2023 were included, with 5 from the United States (Kar et al [29-31], Zhang et al [32], Edalati et al [33]), 1 from China (Shen et al [34]), and 1 from Taiwan (Chang et al [35]). Of these, 6 studies involved patients with breast cancer with additional cancers (eg, sarcoma, lymphoma, leukemia) in some cohorts. Imaging modalities included MRI (4 studies: 3 displacement encoding with stimulated echoes [DENSE] MRI, 1 CMR), echocardiography (n=2), and nongated, noncontrast chest computed tomography (CT) (n=1). AI approaches varied: 4 studies (57%) used convolutional neural networks (CNNs), 1 (14%) used ML, and 6 (86%) implemented image segmentation. Table 1 provides demographic and descriptive data and Table 2 provides details of the AI components of the included studies.

The included studies revealed significant clinical heterogeneity across the studies. Study designs ranged from retrospective (eg, Shen et al [34]: n=1468; Zhang et al [32]: n=260) to prospective (eg, Chang et al [35]: n=211; Kar et al [29-31]: n=32‐42), impacting sample size and follow-up duration (eg, 3 years in Chang et al [35] vs 10 years in Zhang et al [32]). Imaging modalities differed in application: echocardiography (Chang et al [35], Zhang et al [32]) assessed LVEF and GLS; DENSE MRI (Kar et al [29-31]) focused on strain analysis; CT (Shen et al [34]) targeted coronary artery calcium scoring (CACS); and CMR (Edalati et al [33]) evaluated image quality and efficiency. AI techniques showed varied sophistication—CNNs (eg, DeepLabV3+ in Kar et al [29-31], CNN pipeline in Zhang et al [32]) and deep learning (Shen et al [34]) enhanced segmentation and classification, while ML with multilayer perceptron (Chang et al [35]) predicted outcomes like heart failure with reduced ejection fraction.

Outcomes centered on CTR-CVT, with cancer therapy–related cardiac dysfunction assessed in 5 studies (Shen et al [34], Chang et al [35], Kar et al [31], Zhang et al [32], Edalati et al [33]), GLS in 3 (Kar et al [29-31]), and LVEF in 3 (Kar et al [29], Edalati et al [33], Zhang et al [32]). Shen et al [34] uniquely linked CACS to major adverse cardiovascular events (MACE). At the same time, Edalati et al [33] emphasized scan time and signal-to-noise ratio. AI improved detection accuracy (eg, automated CACS in Shen et al [34], GLS computation in Kar et al [30,31]) and efficiency (eg, EasyScan in Edalati et al [33]) compared to manual methods. However, direct comparisons across studies were limited by outcome diversity.

Common limitations included small sample sizes (eg, Chang et al [35], Edalati et al [33]), single-center designs (eg, Kar et al [31], Chang et al [35]), and lack of external validation (eg, Kar et al [31]). Geographic restriction (Shen et al [34], Chinese patients) and technical challenges (eg, segmentation issues in Zhang et al [32]) further constrained generalizability.

The risk of bias assessment began with general considerations for all studies, which included establishing a minimal set of confounders identified by the reviewers as likely to introduce bias in the observed associations. Next, each study was described individually within the framework of an ideal target trial. The consensus results from the evaluations of the 7 nonrandomized studies are depicted in the “traffic light” plot shown in Figure 4.

In 2018, Zhang et al [32] published their work on automating echocardiographic cardiac images using 14,035 echocardiograms collected retrospectively spanning 10 years. Their study included 152 patients diagnosed with CTR-CVT and other patients with other heart conditions such as hypertrophic cardiomyopathy (n=260), amyloidosis (n=81), and pulmonary arterial hypertension [32]. Zhang developed a model for view classification in just a few steps. First, they taught the machine to recognize individual echocardiographic views, where models were trained using manual labels assigned to individual images. Then, they used deep learning architecture for view classification, designed to mimic how the visual system works [32]. This process refers to multiple layers of neurons, processing nodes tuned to recognize features within an image. Afterward, they trained a 13-layer CNN and assessed the accuracy using 5-fold cross-validation. Finally, they used t-distributed stochastic neighbor embedding (an algorithm for visualizing high-dimensional data) to cluster the output of the top layer to visualize the output of their view classification network [32]. By training the CNNs, Zhang could perform image segmentation to locate cardiac chambers that derived cardiac structure and function measurements to develop disease classification models [32]. Zhang’s approach is intended to enable data mining and knowledge extraction from the enormous number of archived echocardiograms, which will have a significant clinical impact by introducing relatively low-cost quantitative metrics into clinical practice and enabling causal insights that require systematic longitudinal tracking of patients [32]. The study results favored using AI-automated measurements over manual measurements across 11 internal consistency metrics. One of these is the correlation between left atrial and left ventricular volumes. This work is argued to have laid the basis for using automated interpretation to support serial patient tracking. Limitations to the study are the length of the analysis period and room for bias. Moreover, the study did not include the number of males or females involved, which may affect the results.

Using a different imaging modality, Edalati et al [33] developed EasyScan, which is automated cardiac planning, by developing, training, and validating 2 deep neural networks on preacquired cardiac MRI datasets (also known as cardiovascular magnetic resonance). EasyScan is implemented with the CMR scanner for automatic slice planning and shimming. The trial included 10 healthy individuals (5 males and 5 females) and 10 cardio-oncology patients (5 males and 5 females) undergoing 2 identical CMR protocols (manual cardiac planning versus AI-based EasyScan) to assess the time difference and accuracy of the cardiac plane. Moreover, Cine images were obtained for the study participants with standard cardiac volume shim and AI-shim to assess the signal-to-noise ratio, contrast-to-noise ratio, overall IQ (sharpness and magnetic resonance image degradation), LVEF, and absolute wall thickening [33]. EasyScan demonstrated accelerated cardiac exams compared to standard manual cardiac planning and achieved an improved and more uniform B0 magnetic field homogeneity using the AI-shim technique compared to volume shimming [33]. Eldalati argued that his results suggest many potential positive outcomes of implementing AI, including a more straightforward and faster workflow chain by minimizing technique complexity. However, a significant limitation of this study is the cohort size, as it is considered small compared to other papers in this field.

Kar et al [29-31] used AI, deep learning, segmentation, and fully convolutional networks (FCN) on the DENSE MRI sequence imaging modality in their 3 studies. In the study published in 2021, Kar et al [29] investigated the automation of measuring left-ventricular strain with a quantification tool via segmentation with a supervised deep convolutional neural network (DCNN) before strain analysis with DENSE images [29]. Kar and her team were able to introduce a novel and automated DCNN architecture–based chamber quantification methodology for detecting the extent of left-ventricular myocardium in single-scan DENSE MRI for patients with breast cancer susceptible to cardiotoxicity. Kar et al identified accurate segmentation, chamber quantification, and subsequent strain analysis in the myocardium as the main critical requirements for engineering and developing this solution. After validation, Kar et al emphasized that their DCNN-based segmentation can provide accurate estimates of the left-ventricular chamber quantification required in strain analysis.

Kar et al argued that their model can perform fast and inexpensive automated measurements of cardiac strain as the model can detect altered material properties. However, the thresholds that define cardiac dysfunction caused by cancer therapy are still an area that needs to be further studied [29].

In 2022, Kar and her team continued their work using DENSE in developing another direct MRI-based, FCN-based, deep-learning semantic segmentation approach for computing GLS for patients with breast cancer [30]. This time, they computed myocardial strains directly from the unwrapped phases with the radial point interpolation method. They compared the results of 30 patients with 30 healthy individuals, and the difference in GLS results between the participants demonstrated that the FCN is sensitive to unwrapping left ventricular data in a heterogeneous cohort [30]. Moving forward with their work on GLS computation, Kar and her team investigated early alterations in prognostic factors such as GLS with standard Cox proportional hazards regression for estimating the risk of CTR-CVT incidents in patients with breast cancer undergoing cancer treatment using their previously developed AI-FCN.

Moving forward, Kar and her team carried out a trial using their tool to estimate the risk of developing cardiotoxicity in patients with breast cancer using data from their previous studies [31]. The trial proved their hypothesis that GLS computation can be used for early detection of CTR-CVT as an independent prognostic method of left ventricular dysfunction [31]. The advantage Kar et al had in their studies was that they were able to validate their solution internally within their center. However, their trials did not come without limitations. The solutions were not validated externally with other centers, and there was a greater sample for better accuracy measures [29-31]. In addition, the phase unwrapping approach for GLS measures was not compared to phase wrapping with another FCN, such as PhaseNet, which is considered a significant limitation in their conclusion [30].

Concurrently, in 2022, Chang et al [35] conducted another single-center prospective study and included a larger sample size of 211 patients diagnosed with breast cancer at different stages [35]. Chang et al [35] aimed to establish an AI-based predictive model for CTR-CVT using a cardio-oncology program. They prospectively collected clinical information and echocardiographic images from patients with breast cancer over 1 year. In their study, 2 echo technicians performed an echocardiogram independently to measure the LVEF at baseline, 3 months, 6 months, and 1 year after patients received their treatment. A cardiologist with a validated reliability and reproducibility interpreted the images. Moving forward with the AI solution, data were validated using a data mart for further analysis. Then, we compared the accuracy, precision, sensitivity, specificity, and area under the curve of the random forest, logistic regression, support vector clustering, LightGBM, K-nearest neighbour, and multilayer perceptron models. This process yielded the best accuracy in predicting CTR-CVT [35]. Moreover, the multilayer perceptron showed the best results in predicting heart failure with a reduced ejection fraction as an early sign of myocardial dysfunction after the occurrence of CTR-CVT [35].

Shen et al [34] conducted the most recent study in China in 2023. The study aimed to evaluate whether the pretreatment CACS can stratify the risk of CTR-CVT and MACEs in patients with diffuse large B-cell lymphoma (DLBCL). They retrospectively collected nongated and noncontrast chest CT scans of 1468 patients from 4 health centers in China, then used a deep-learning–based algorithm software (CACScoreDoc) to calculate the automatic CACS. CACScoreDoc automatically calculated the CACS and transmitted the results to the doctors after uploading the CT images to the software. The study showed that automating CACS derived from chest CT scans done before receiving the treatment is potentially helpful in identifying patients at risk of developing CTR-CVT and MACEs in patients with DLBCL receiving anthracycline chemotherapy, which can guide clinicians to implement cardiovascular protective strategies and minimize CTR-CVT in DLBCL patients [34].

Although cardiovascular events that are caused by cancer medications vary in prevalence from one type of cancer and its medication to another, they are still the second most common cause of mortality in cancer survivors. To accurately predict the risk of cardiotoxicity among individuals receiving cancer treatment is still a great challenge in the cardio-oncology field due to high cost, limited access to care, and inadequate compliance with screening protocols. Therefore, noninvasive, low-cost, accessible, innovative approaches to predict high-risk individuals and detect cardiotoxicity early among patients with cancer are critically needed to enable optimal screening, early diagnosis, and timely interventions [36].

The current tool used to investigate signs of cardiotoxicity is medical imaging, with the 2 most used imaging modalities for this purpose being the echocardiograph and CMR. However, although these modalities have helped the medical field to achieve significant improvement in prognosis in this area, some drawbacks hold them back from being optimal methods of investigation. The echocardiograph is entirely user-dependent in image reproducibility and results interpretation, leaving ample room for bias and inconsistency. On the other hand, the CMR is not always available due to its high cost. Therefore, more robust, cost-effective methods and imaging protocols are needed in this cardio-oncology area to optimize patient care [36].

Many health care disciplines have moved toward advancing artificial intelligence and developing better ML algorithms as they continue to improve patient care quality significantly. With the availability of enormous volumes of patient data and accessibility of proper hardware, AI and ML can accelerate the pace of change in health care. These technologies can sift through the data and analyze it much faster than humans, leading to increased efficiency. ML is used to predict clinical risk factors by feeding it with an enormous volume of data retrieved from patient medical records or national datasets and registries or detect cardiotoxicity via deep learning of patients’ cardiovascular images. In this review, the authors focused their assessment on using AI and ML in cardiovascular imaging to increase the diagnostic strength and accuracy in detecting CTR-CVT.

This review included 7 studies that intended to assess the implementation of AI in cardiovascular imaging among patients with cancer. These studies examine the use of AI on MRI, echocardiogram, and CT imaging modalities with different AI technologies such as ML, CNNs, and image segmentation.

The future of imaging AI in cardio-oncology holds substantial promise. This convergence of cutting-edge technologies, encompassing molecular imaging, wearable devices, multiomics data, and predictive modeling, is poised to transform cardiotoxicity management in patients with cancer. These advancements enable early detection and personalized risk assessment and promise targeted interventions, ultimately enhancing patient outcomes and survivorship. This future trajectory in imaging AI aligns with the significant advancements witnessed from ML to deep learning in AI, revolutionizing robotics and autonomous systems’ capabilities and enabling them to perceive, learn, and adapt with increased efficiency and accuracy in complex environments. These models, leveraging AI algorithms trained on diverse patient cohorts and multimodal imaging data, could assist clinicians in formulating proactive strategies for long-term cardiac care in cancer survivors, thereby enhancing overall cardiovascular health and quality of life.

As promising as AI and ML sound to the advancement of imaging in health care and the prediction of the risk of developing cardiotoxicity among patients receiving cancer treatment specifically, there are methodological and practical limitations preventing these technologies from reaching their full potential. The evidence base needs more prospective validation of the technology and current workflow, including evidence on the length of analysis required for validation and the interoperator and interobserver variability to eliminate manufactured variations that limit reproducibility [23]. Moreover, their usefulness in health care depends on incorporating the AI tool in clinical decision-making as part of the clinical practice routine, and that concern needs further investigation [37]. Another inadequacy of AI applications in health care is the systematic biases affecting patient demographics, such as gender imbalance [38]. It is worth mentioning that AI requires training on all kinds of populations with different demographics to guarantee equal performance from one population to another. It is recommended that multiple massive datasets be combined either retrospectively or prospectively to improve the generalizability of the ML process and the training of AI models, which was not achieved by all the included studies in this review [39].

The first limitation we had while conducting this review was the limited published evidence in the literature about the application of imaging AI in cardio-oncology to predict CTR-CVT. Therefore, we could not specify the cancer type or treatment under investigation. Second, even though there is significant literature on AI and imaging with different modalities, when we narrowed it down to our criteria, which was patients with cancer who are undergoing cardiotoxicity assessment, the literature search resulted in 3 different imaging modalities rather than studying AI with one specific imaging technique at a time. This resulted in different outcomes that prevented us from proceeding with a meta-analysis.

The use of AI in the medical field is a relatively new research area. This review could be used to stimulate further research. It can be used as groundwork for lab work to improve AI models or inspire new ones. In addition, this review highlights the positive outcomes of different studies in this area and their limitations. It may encourage experts to improve the AI and ML models and eventually implement them into medical imaging, possibly leading to the advancement of the field. However, given this field’s rapidly evolving nature, additional studies may have been published since the initial search process for this paper.

In conclusion, this systematic review highlights the promising potential of AI in enhancing cardio-oncology imaging for predicting cardiotoxicity in patients with cancer. Through analyzing 7 studies conducted between 2018 and 2023, it became evident that AI methodologies, including ML and deep learning, can significantly improve the accuracy and efficiency of cardiotoxicity assessments across various imaging modalities, such as echocardiography and CMR.

The review underscores that AI-driven tools have demonstrated improved clinical outcomes by enabling earlier detection of cardiovascular complications associated with cancer therapies. However, while the findings are encouraging, the limited number of studies and their varying methodologies indicate a need for further research. This includes conducting larger, multicenter trials to validate AI applications in diverse patient populations and refine these technologies for routine clinical use.

In light of these insights, collaboration among data scientists, health care professionals, and researchers is essential to advancing AI’s integration in cardio-oncology. This collaboration will pave the way for personalized medicine approaches, ultimately enhancing patient care and improving the quality of life for cancer survivors at risk of cardiotoxicity.