Introduction

In the upcoming decades, it is anticipated that cancer would surpass other illnesses as one of the main global causes of morbidity and mortality [1]. A recent study from The Lancet [2] demonstrated that for many low-income and middle-income nations, noncommunicable diseases (NCDs) pose an ever-greater health threat, with cancer becoming an NCD of greater importance. Therefore, it is imperative to focus on cancer treatment, enhance the rate of early detection and cure, and boost cancer screening.

Due to technical advancements in statistics and computer software, computer professionals, and health scientists may now collaborate closely to improve prognoses. As a result of the adoption of artificial intelligence (AI) strategies, researchers have increasingly concentrated on creating models using AI algorithms to detect and diagnose cancer. AI is the process of teaching a computer to mimic human intelligence by showing it how to study, evaluate, comprehend, deduce, interact, and make decisions [3]. Tremendous success has been achieved with AI in the last ten years in the fields of speech synthesis, natural language processing, and computer vision. This review focuses on the latest AI techniques for tumor diagnosis, treatment, and prognosis. We highlight artificial intelligence applications for omics data types, and we offer perspectives on how the various data types might be combined to create decision-support tools and discuss how current challenges may be surmounted to make artificial intelligence useful in clinical settings in the future.

We searched three databases from their creation until November 10, 2023: MEDLINE (PubMed), CENTRAL (Cochrane Central Register of Controlled Trials), and Embase to assess the published literature pertaining to the application of artificial intelligence in cancer. Due to the rapid pace of AI updates, we have focused on the last two years of relevant research. The following keywords were used in this sco** review: (neoplasms OR cancer) AND (artificial intelligence OR deep learning OR machine learning). With a focus on the application and usage of artificial intelligence in cancer treatment, we incorporated a total of 254 publications in the construction of this narrative review, including pertinent prospective, retrospective, and review studies.

Specific meaning of artificial intelligence

AI is an area of computer technology comprising numerous techniques and subfields aimed at performing activities that could previously be completed only by humans. To enhance the interpretation of medical data relevant to medical administration, diagnostics, and predictive outcomes, AI technologies and their subdomains are being implemented in healthcare delivery. The two main techniques for implementing AI are machine learning (ML) and deep learning (DL), which are terms that are frequently used interchangeably. Deep learning is a branch of machine learning. ML generates predictions by spotting patterns in data by means of mathematical algorithms. DL produces forecasts using multiple layers of fabric neural network algorithms that are modeled after the brain’s neural network architecture. In the past ten years, with advancements in big data, algorithms, computing power, and Internet technology, AI has excelled in numerous tasks across a wide range of industries, including identification of faces, image classification, speech recognition, automatic translation, and healthcare [4]. The main ML techniques are support vector machines (SVMs), decision trees, and K unsupervised algorithms, while the most commonly used for DL today are convolutional neural networks (CNNs) [5]. Figure 1 presents a few of the most basic ML and DL approaches.

Fig. 1
figure 1

Network structure of DL. a A model of an SVM; b A model of a random forest that is composed of several decision trees; c KNN characterized by the fact that it is composed of many random features rather than a linear feature; d Components of CNNs [6]; and e Components of graphical CNNs

Full size image

ML fundamentally seeks to replicate or mimic humans’ capacity for pattern recognition. Traditional ML approaches take far longer to teach and test based on a specific problem than DL approaches. SVMs, decision trees, random forests, gradient boosting (such as XGBoost), and other conventional ML techniques are examples of traditional ML techniques. There is a significant flaw with decision trees, namely a decision tree divides samples extremely precisely, but dividing samples too precisely causes overfitting of the training set, and dividing samples coarsely results in a decision tree that does not fit the samples properly. Decision trees called random forests are based on the concept of learning and bagging combined. Two factors—the random selection of the dataset and the random selection of the characteristics utilized in each tree—reflect the unpredictability of a random forest the most. The XGBoost technique repeatedly constructs an ensemble of decision trees. The capacity of this technique to manage missing data, capture nonlinear correlations between the model features and the outcome, and have higher-order interactions between variables is its key benefits over conventional logistic regression-based risk models [7].

Training artificial intelligence models

Several processes are necessary for training an AI model, including data gathering and preparation, model selection, model training, and hyperparameter tuning.

Data collection and preprocessing

With the rapid development of modern medicine, various types of data are emerging. A large amount of imaging data has been generated as represented by X-ray, CT, and MRI, and the development of pathology has made sectioning the gold standard for tumor diagnosis. In addition to the traditional clinical information data, with the remarkable advances in sequencing technology over the past two decades, how to deal with the large amount of molecular data brought about by genomics, transcriptomics, proteomics, etc., has also become a matter of close attention for clinicians. Later, we will describe how to deal with a single type of data. However, a patient usually does not have only one type of test and one type of data, so we will also introduce how to integrate different types of data to enhance computational models.

To facilitate the subsequent model training, we need to preprocess these data. For digital data, we need to remove outliers, deal with missing values, and normalize the data. The most often utilized AI algorithms, using EHR as an example, are deep learning, decision tree algorithms, and regression algorithms. While completing regression tasks to finish disease risk prediction, researchers also use classification tasks to extract lesion characteristics from illnesses and categorize them [45, 46]. A DL system [47] was created that can correctly identify the existence of lung cancer within three years and account for all pertinent nodule and nonnodule markers on screening chest CTs. Their research was the first to create a deep machine learning prediction method without the use of computer-aided diagnostic tools to assess a person’s 3-year probability of develo** lung cancer and related lung cancer-specific mortality. Kiran Vaidhya Venkadesh et al. [48] created and externally verified a CNN-based DL algorithm for estimating the likelihood of malignancy in lung nodules found by low-dose screening CT, which demonstrated good performance, on par with thoracic radiologists, at estimating the malignancy risk of pulmonary nodules observed during screening CT(AUC = 0.93). However, their researches included a number of restrictions. Firstly, one CT scan was employed in the created method, and a prior CT image was not taken into account [48]. Secondly, on average, members of the cohort [47] had LDCTs for screening every year, which may cause bias in the measurement results.

In order to address the above issues, a deep learning system was developed that can forecast the probability of develo** lung cancer six years from now. Newly developed Sybil [90] to classify skin lesions and to present dermatologist-level prediction outcomes. The knowledge distillation approach is also often used to help diagnose melanoma [91, 92]. In addition to the simple teacher–student model, the SSD-KD approach [93], a unique self-supervised diversified knowledge distillation technique, has been used for the lightweight multiclass categorization of skin diseases utilizing dermoscopy images. In that study, the conventional single relational modeling block was substituted with dual relational blocks in terms of technological innovation. Multi-Site Cross-Organ Calibrated Deep Learning (MuSClD), a novel approach to cross-organ calibration between two sites of digitalized histopathology images, was validated in nonmelanoma skin cancer. 3D images [94, 95], EfficientNet [96, 97], genetic programming (GP) [98], and new AI algorithms on smartphones [99, 100] have also been developed for skin cancer diagnosis.

To supplement human visual inspection, AI can assist in the detection of undetectable tumor lesions on PET scans. Ga-PSMA-11 PET-based radiomics features have been used to generate random forest models that accurately predicted invisible intraprostatic lesions [101]. Biopsy and magnetic resonance imaging (MRI) are frequently used to diagnose intracranial tumors. Due to the similar phenotypes of various tumor classes on MRI scans, it has been difficult to identify tumor types, especially rare types, from MRI data. A DL method for segmenting and classifying 18 distinct types of intracranial tumors was developed [102] using T1- and T2-weighted images and T2 contrast MRI sequences and evaluated with an AUC of 0.92.

AI may easily be applied to medical imaging, and major advancements in this area have been made in recent years. AI eliminates the uncertainty that people contribute to decisions and delivers objective measurements for each choice. However, the limits are also readily apparent. The molecular causes of illnesses are not revealed by morphological evidence. By using this method, disease states with the same morphological appearance cannot be discriminated.

Tumor staging and grading

Important factors for tumor T-staging include the size and degree of invasiveness of primary tumors, which comprise descriptions of their shapes. Convolutional neural networks are most used in this task. The T stage of Barrett’s carcinoma is a crucial consideration when choosing a course of therapy. Endoscopic ultrasonography is still the norm for preoperative staging, but its usefulness is under question. To help with staging and to improve outcomes, new tools are needed. With a high accuracy of 73% in diagnosing esophageal cancer, an AI system built around endoscopic images has been developed [103]. Tumor sizes and forms vary, making individual slice-by-slice screening for T-staging time intensive. Consequently, a multi-perspective aggregation network (TSD Net) has been created with ideas from oncological diagnostics that included different diagnosis-oriented knowledge and enabled automatic nasopharyngeal carcinoma T-staging identification [104].

Advances in imaging histology have greatly contributed to hel** TNM staging of tumors. Separate iterations of the machine learning models have been created using both the entire collection of extracted features (full model) and just a selection of the previously discovered robust metrics (robust models) to confirm that CT-based radiomics signatures were effective tools for determining the grade and stage of ccRCC [105]. Additionally important in the early phases of decision-making, but time-consuming, is a delineation of the tumor. To forecast the grade of a tumor while also segmenting it, a single multi-task convolutional neural network has been created using the whole 3D, structural, preoperative MRI data [106].

Accurate assessment of lymph node metastasis (LNM) is essential for evaluating the staging and grading of tumor patients. In addition to offering a straightforward “yes” or “no” response on the likelihood of having cancer, AI models can also identify the disease site from a test picture. One of the most common applications is to help find the localization of metastatic tumors. Using whole-body PET/CT scans, convolutional neural networks (CNNs) based on UNet [1] were trained to detect and separate metastatic prostate cancer lesions fully automatically [107]. The localization of tumor metastasis in whole-slide images has also been studied extensively in recent years [107,108,109,110]. The condition of the lymph nodes (LNs) prior to surgery is crucial for the management of colorectal cancer (CRC). With areas under the curve (AUCs) of 0.79, 0.73, and 0.70 in the training set, testing set, and verification set, respectively, a deep learning (DL) model [111] with features gathered from improved venous-phase CT images of CRC has been proposed to identify LNM in CRC. Shaoxu Wu et al. [112] created a diagnostic algorithm called LNMDM based on AI that was effective for finding micrometastases in lymph nodes and was demonstrated not only in bladder cancer (0·983 [95% CI 0·941–0·998]) but also in breast cancer (0·943 [95% CI 0·918–0·969]) and prostate cancer (0·922 [95% CI 0·884–0·960]). AI plays a significant role in aiding diagnostics to find lymph node metastases in slide pictures. Lymph node metastases, especially micrometastases, were successfully identified by the LNMDM [112] on whole-slide images in bladder cancer. The VIS AI algorithm demonstrated comparable accuracy and NPV in identifying LN metastases on breast cancer. In summary, the implementation of AI in tumor staging and grading has significantly improved tumor prognoses and increased the general survival rate of cancer patients.

Tumor therapy

AI for exploring tumor therapeutic targets

In recent years, the development of multiomics technologies in cancer research [113, 114] has greatly facilitated the discovery of anticancer targets [115,116,117]. The advancement of precision medicine and translational medicine will be significantly aided by the use of ML and DL to mine multiomics data to investigate complicated disease causation processes and treatment response mechanisms. In the following, we describe in detail the advances in genomics, epigenetics, transcriptomics, proteomics, metabolomics, and multiomics in cancer target discovery. Figure 3 describes the main sources of these six components and the advanced methods currently comprising them.

Fig. 3
figure 3

Components of multiomics and the main techniques. The combination of AI and multiomics has led to the discovery of new targets for cancer therapy. (Created with BioRender.com)

Full size image

Genomics

The genome contains inherited information that controls gene expression to shape the structure and working machinery of the cell [118]. Genomics focuses on understanding the composition, organization, visualization, and modification of an organism’s whole genome [119]. The rise of the genomic era has also boosted precision medicine and cancer [120]. The approach of a meta-learning model [121] allows users to discover significant pathways in cancer and priority genes based on their contribution to survival prediction. To fully understand how cancer develops, progresses, and is treated, accurate somatic mutation detection is difficult yet essential. The first method for detecting somatic mutations based on deep CNNs is called NeuSomatic [122]. However, the fact that matched normal specimens are not frequently acquired in clinical practice is a major barrier to genetic testing in cancer. The somatic vs. germline status of each discovered change may be predicted using SGZ, [123] which does not need a patient-related standard control, by modeling the mutation’s allele frequency (AF), accounting for the cancer content, cancer ploidy, and local copy number. Similarly, a recently created method, Continuous Representation of Codon Switches [124] (CRCS), a DL-based technique, can aid in the identification and investigation of driver genes as well as the detection of cancer-related somatic mutations in the absence of matched normal samples.

Taking colon cancer as an example, numerous studies [125,126,127,128] have subtyped colorectal cancer based on similar and different biological traits and pathways, and they have identified the relationships between these pathways and patient prognosis, overall survival, and responsiveness to various treatments—particularly targeted therapy and immunotherapy. Using 499 primary colorectal neoplasm diagnostic images from 502 individuals in The Cancer Genome Atlas Colon and Rectal Cancer (TCGA-CRC-DX) cohort, a retrospective study established a weakly supervised DL framework incorporating three separate CNN models [85]. After comprehensive validation, the method was shown to be helpful for patient classification for targeted medicines, with possible cost savings and quicker turnaround times compared to sequencing- or immunohistochemistry-based techniques. The research, however, examined each individual image tile without considering the significance of the spatial relationship between tiles. In a recent study [129], a method for forecasting cross-level molecular profiles involving gene mutations, copy number variations, and functional protein expression from whole-slide pictures was proposed. This method focuses on the spatialization of cancer tiles. In the training dataset, the model performed exceptionally well in predicting a variety of genetic alterations and then identifying targeted therapies for colon cancer patients.

Epigenetics

Epigenetic modification is the genetic change in the way genes operate and express without altering the DNA sequence. DNA methylation, histone modification, and chromatin structure manipulation are the three primary epigenetic modifications that are now understood [130]. Although there are high-quality data on DNA methylation, few samples have RNA-seq data due to numerous experimental difficulties. Therefore, an innovative technique called TDimpute [131] was created to reconstruct lost data on gene expression from DNA methylation data using a transfer learning-based neural network. Understanding how epigenetics regulates gene expression to govern cell functional heterogeneity is dependent on the ability to predict differentially expressed genes (DEGs) from epigenetic signal information. On the basis of epigenetic data, a multiple self-attention model (Epi-MSA) [132] was suggested to predict DEGs. To determine which gene locations are crucial for forecasting DEGs, Epi-MSA first applies CNNs for neighborhood bin information embedding and then makes use of several self-attention encoders on various input epigenetic parameters.

Transcriptomics

Transcriptomics is a useful tool for comprehending the physiology of cancer and locating biomarkers. It includes analyses of alternative transcription and alternative polyadenylation, detection of integration transcripts, investigations of noncoding RNAs, transcript annotation, and finding novel transcripts [133]. One study using DL algorithms to interpret common cancer transcriptome markers [134] showed that across a wide range of solid tumor types, dysregulation of RNA-processing genes and aberrant splicing are widespread traits on which fundamental cancer pathways may converge. Molecular pathology plays an important role in cancer, but whether it is possible to estimate the levels of gene expression based on a visual inspection of H&E-stained WSIs has never been thoroughly explored. Numerous studies have been conducted to predict cancer gene expression, including that of prostate [135] and breast [136] cancers, across the transcriptome from histopathological images. A DL model called HE2RNA [137] based on a multitasking poorly supervised technique was created using matched WSIs and RNA-Seq profiles from TCGA data, which included 8725 patients and 28 distinct cancer types. This increases the likelihood of discovering novel gene targets. Patients’ responses to treatment are significantly influenced by the quantity, composition, and geographic distribution of the cell groups in the tumor microenvironment (TME) [138]. The thorough characterization of gene regulation in the TME has been made possible by recent developments in spatial transcriptomics (ST) [139, 140]. Three new approaches have recently been developed: Kassandra [141], XFuse [142], and TESLA [143]. Kassandra is a tree ML algorithm that was taught to precisely rebuild the tumor microenvironment (TME) using a large database of > 9,400 tissue- and blood-sorted cell RNA profiles combined into millions of artificial transcriptomes. According to Kassandra’s deconvolution of TME components, these populations play a part in tumor etiology and other biological processes. By utilizing data from H&E-stained histological images, XFuse predicts superresolution gene expression per pixel. TESLA is an ML framework that incorporates gene expression and histological image data into ST to study the TME. The innovative aspect of TESLA is the annotation of diverse immune and tumor cells on histological images directly.

In addition, the identification of lncRNAs [144,145,146] and microRNAs [147, 148] by ML can assist in the precise treatment of cancer. In the fight against cancer, therapeutic decisions are increasingly based on molecular tumor features, and cancer tissue molecular profiling is becoming an essential component of standard diagnosis [149]. To reduce individualized patient differences, scGeneRAI [150] uses layerwise relevance propagation (LRP), an explainable AI technique, to extrapolate individual cell gene regulation networks from single-cell RNA sequencing data. Oncology drug response is a major challenge in cancer treatment. With an average Matthew correlation coefficient (MCC) and AUC of 0.56 and 0.80, respectively, the classification and regression tree (CART) model from interpretable ML models has proven to be the best model for predicting how breast cancer would react to doxorubicin [151]. At the single-cell level, ScDEAL is a deep transfer learning system that integrates bulk cell-line data to predict cancer medication response at the single-cell level. Finding drug resistance targets at the level of transcriptional profiles using AI deserves more research in the future.

Proteomics

Proteomics is a broad study of proteins that identifies and counts the proteins present in a biological sample, such as a sample of cells, tissues, or bodily fluids. Proteomics data offer the benefit of providing a numerical number of individual proteins throughout the body and dynamic characteristics that develop over time and among individual subjects, in contrast to other forms of omics data, such as genomic data. Mass spectrometry (MS) is a key tool used in proteomics research [125]. MS-based proteomics has advanced quickly in terms of lower cost and higher throughput, regularly permitting large-cohort studies with tens of thousands of participants and tens of millions of identified proteins in cancer cells and other biological samples. However, the majority of research concentrates on the final proteins discovered using a collection of algorithms that compare partial MS spectra with the ordered database, leaving the problem of pattern identification and categorization of the raw mass-spectrometric information unanswered. Consequently, for the analysis of massive MS data using deep neural networks (DNNs), the publicly available MSpectraAI [152] platform and the tumor classifier [153] have been developed, which could expand the intriguing use of DL techniques for classifying and predicting proteomics data from multiple cancer types and distinguishing between tumor and nontumor samples.

Sequential Window Acquisition of all Theoretical Mass Spectra-MS (SWATH-MS) is a cutting-edge MS method that enables the measurement of nearly all peptides as well as proteins present in a single sample, making it valuable in research involving massive sample cohorts [154]. It can be used to facilitate the categorization of CRC molecular subgroups and promote both diagnostics and the creation of novel medications [155]. Regarding colorectal cancer, a mechanism-based ML approach [156] has been proposed to find genes and proteins with substantial correlations to event-free patient survival and predictive potential to account for patient-specific variations in STN activity by building three linear regression models. The development of proteomics has contributed to the discovery of new targets in hematological tumors. Targetable enzyme characteristics have been revealed by proteomics of acute lymphoblastic leukemia that is resistant to Notch1 suppression [157]. Through the induction of long-lasting immune responses, T cells play critical roles in human defense against hematological tumors. In recent work [158], ML and nanoscale proteomics were coupled to subtype T cells in peripheral bloodstreams from single individuals with multiple myeloma. To reduce the possibility of overfitting the ML models, differentially expressed proteins (DEPs) were selected according to statistical significance, and only the top 13–15 DEPs were utilized. Thus, this work helped identify new targets for immunotherapy. Another DL network [159] identified the 20 proteins most strongly associated with FLT3-ITD in acute myeloid leukemia. In addition, DL and ML have been applied to proteomics data for pancreatic cancer [160] and diffuse large B-cell lymphoma [161] patients, respectively.

Metabolomics

Metabolomics is a burgeoning area of research that utilizes technologically sophisticated analytical chemistry to perform high-throughput characterization of metabolites in cells, organs, tissues, or biological fluids [162]. New therapeutic targets have been suggested to target metabolic constraints in cancer as a result of metabolomics studies, which have revealed potential medicinal weak points for treating cancer [163]. Lipidomics is a branch of metabolomics that aims to study and analyze the lipids in the metabolome and the molecules that interact with them [164]. Metabolomics analysis can be performed using GC‒MS and LC‒MS, and LC‒MS is commonly used for the analysis of lipidomics. The combination of metabolomics and AI has flourished in various areas of cancer, including breast cancer [165, 166], head and neck cancer [167], colorectal cancer [168, 169], glioma cancer [170], esophageal cancer [171, 172], lung cancer [249]. The most susceptible source of information determines the total security level when we combine patient data from other sources. Clinical data are frequently the property of particular institutions due to concerns about patient privacy, and there are few methods in place to share data among institutions. It is frequently inadequate to remove personal identifiers and secret information since an attacker can still draw conclusions to retrieve some of the missing data. The good news is that multicenter information transfer agreements and safeguarding privacy distributed DL (DDL) are starting to overcome this roadblock [250,251,252]. DDL offers a mechanism that protects privacy so that several users can collaborate on learning using a deep model without directly exchanging local datasets. In addition, it is important to ascertain the level of supervision that doctors must provide and identify the person accountable for any poor choices made by DL tools. On the other hand, we should educate AI users to guarantee that they are knowledgeable consumers of the technology and endeavor to openly and clearly express to them what they should anticipate in a variety of circumstances. Many of the hazards described above may be reduced by being accessible, having varied demands, and being cautious.

When implementing AI, ethical issues are crucial since unethical data gathering or usage practices might introduce biases into models. These biases can take numerous forms, but they are mostly determined by the data and cohort composition employed by the particular AI systems. Providing and reviewing AI models lacks defined criteria or norms. Identifying the possible biases included in the established systems will be crucial; thus, future studies should fill this knowledge gap to help researchers and physicians.

Clinical integration

As mentioned above, AI has been shown in many studies to improve the correctness of cancer diagnosis. However, a different perspective has been proposed. In one study [253], the authors systematically evaluated 131 published studies using the QUality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2) tool. They reported that the accuracy of AI in breast cancer screening programs cannot currently be evaluated based on available research, and it is unclear where in the therapeutic pathway AI could be most helpful.

For LLMs, one of their advantages is the capacity to sift through vast volumes of data and provide replies in a conversational and understandable manner. LLMs also have the potential to be used in patient education and consultation, offering patient-friendly information to aid in their understanding of their medical issues and available treatment choices, facilitating joint decision-making. More crucially, LLMs can contribute to the democratization of medical knowledge by allowing anybody, regardless of location or socioeconomic position, quick access to reliable medical information. However, special attention needs to be paid to the fact that current LLMs are not yet capable of fully replacing doctors, as they may contain errors or omit key points in the responses. Although ChatGPT-4.0 was more accurate than the other tools, neither ChatGPT nor Google Bard or the Bing or Google search engines provided 100% accurate answers to all queries [242]. The much-anticipated Med-PaLM, while promising, is evaluated by multiple choice questions; however, real life is not multiple choice, and different clinical symptoms and specificities in different patients make clinical diagnosis more complex. While AI, such as ChatGPT-4.0, might be helpful for giving broad information and responding to frequently asked queries. Nonetheless, it is important to take great caution when responding to inquiries from certain patients. It is essential to continuously upgrade AI models to include the most recent medical information.

Currently, almost all relevant AI models have been created to assist in cancer diagnosis using clinical data from the time of development. These clinical data may be derived from patient reports, complaints, or sequencing results. The question is whether there is an AI model that can recommend more tests and treatment modalities or perhaps aid in prescribing anticancer medication without relying on clinical data. The current state of affairs is that with the development of multiomics, a variety of data, such as methylation and fragmentomics [254], are being used to train AI models. If one day the data of the AI model accumulates to a large enough size, is it possible to predict the probability of cancer occurrence by only entering the data of normal people, and is it possible to give the corresponding chemotherapy regimen by only comparing the sequencing results of cancer patients and the database. This is a question worth thinking about and very interesting. First, the database must be large enough and ethical; second, there is variability between individuals, and it would be irresponsible to treat them by looking only at sequencing data at the genetic level or transcriptional level, for example.

However, if it is only in the area of cancer diagnosis, AI models have the potential to identify molecules and biomarkers associated with mutated genes and thus confirm the diagnosis of cancer independently of traditional pathology measurements. Meanwhile, with the advent of wearable and portable medical instruments, AI has shown much potential for the early screening of tumors. Therefore, we think that in the future, AI models have the potential to impact the cancer diagnostic market, but in terms of treatment, they cannot be separated from doctors and clinical data.

What must be realized is that despite the rapid development and promising future of AI, it can never replace clinicians and will only become an important tool to assist them in the future.

Conclusion

In summary, AI has the ability to fundamentally alter cancer treatment and move it closer to the promise of precision oncology. In an era where genomics is being incorporated into health delivery and health data are being more digitized, it is anticipated that AI would be used in the construction, verification, and application of decision-support tools to promote precision oncology. We highlighted several promising AI applications in this review, including detection, prognosis, and administration of cancer treatments. It is undeniable that large language model can greatly assist physicians in their clinical work, but it can never replace them. Important conditions for the general adoption of AI in clinical settings include phenotypically rich data for the development of models and clinical validation of the biological value of AI-generated insights. Finally, clinical validation of AI is required before it may be used in ordinary patient treatment.