It was estimated that there were >19 million new cancer cases and 9 million cancer-related deaths in 2020 worldwide. The top five cancer sites, based on the global cancer incidence, are as follows: Female breast cancer (BC), lung cancer (LC), colorectal cancer (CRC), prostate cancer (PCa) and stomach cancer, and more than half of cancer deaths globally are attributed to these cancer types (World Health Organization. Cancer Today. 2020; http://gco.iarc.fr/). In China, a growing cancer burden was unexpectedly observed, and nearly 3 million cancer deaths and >4 million individuals were diagnosed with cancer in 2020. Accurate diagnosis and personalized cancer therapy are the keys to improving the cancer prognosis for every patient. With the promise of an estimated numerical prognosis for patients with cancer, AI has been proposed as a prominent way to improve pathological diagnostic accuracy (DA) and cancer prognostication (CP). Over the past decades, although the exploration and exploitation of AI in cancer pathology has evolved substantially, their generations, interpretations and impact on patients have remained incompletely understood by oncologists. In the present study, the published literature on Google Scholar and PubMed were searched using the following terms: ‘AI’ AND ‘pathology’ AND ‘cancer’, or ‘machine learning (ML)’ AND ‘pathology’ AND ‘cancer’, or ‘deep learning’ AND ‘pathology’ AND ‘cancer’. The present review includes an overview of the roles of AI-integrated cancer pathology in precision medicine by summarizing the rationale for its use, clarifying recent innovations and grouping the main applications, as well as discussing the challenges for AI-integrated cancer pathology-based individualized treatment.

AI is defined as the ability of a machine to simulate, extend and expand human cognition, and has the most features suggestive of human intelligence, to decide upon an action to achieve the desired goal. Initially proposed in 1950 by Turing with ‘Can Machines Think’, and subsequently coined in 1956 by McCarthy (Fig. 1), AI was originally described as being able to ‘simulate intelligence’. With the gradual advent of ML based on computational algorithms and the following computational intelligence, the more power of the decision-making pattern AI was endued with, and the more diverse, complex or esoteric fields AI was involved in, a marked impact was made on modern civilization, namely on Industrial Revolution 4.0 (1). Deep learning (DL) introduced in 1986 by Dechter is by far the most common subtype of ML, and represents computational models originating from artificial neural networks (ANNs) to derive progressively higher-order features from data in the form of multiple non-linear layers (Fig. 2) (1). Inspired by neurobiology, DL is composed of units that compute a weighted sum of the multiple inputs (referred to as the pre-activation) and transforms the results non-linearly, which includes automatic learning and hierarchical representation on multiple levels (2). DL has been employed in nearly all scientific areas, especially in medical imaging, including cancer pathology (the cornerstone of cancer medicine), which has yielded important results that paved the way for the development of the convolutional neural network (CNN), generative adversarial network (GAN), auto encode and so on, to meet the demands of specific biomedical applications for precision oncology.

Defined as an image-based approach, DP aims to acquire, manage, interpret and distribute pathology information in a computer-empowered manner, to extract and analyze pathological visual data. Since the introduction of telepathology in 1986 by Weinstein, the emergence of whole slide images (WSIs) in the 1990s, and other digital images (Electronic microscope digital images), has boosted the prosperity of the DP era (Fig. 1A) (3). Efforts have been devoted to improving the scanning speed and the accuracy of images, which are the two major issues affecting the throughout performance of WSIs. Notably, GAN, an unsupervised technique proposed by Goodfellow et al (4) in 2014, was composed of two competing models: i) A generative model G capturing the data distribution; and ii) a discriminative model D, which made likelihood estimations on the training data rather than G. With respect to this, Fanous and Popescu (5) reported a new method in 2022 (referred to as GANscan) to significantly increase the scanning speed of the whole slide scanner and to correct any defocusing by integrating the Pix2Pix image translation framework and GAN-based model.

Notably, WSIs provide an enabling AI platform for generating a rich variety of novel ecosystems in DP, which is known as Pathomics. Pathomics is an AI-based multi-systems integration methodology to understand cellular interactions and signaling by analyzing relevant data from tissues and cells. With the emergence of Pathomics, integration of AI into the DP workflow has achieved some breakthroughs in cancer pathology. For example, in MICCAI 2014 (Table I), the brain tumor digital pathology challenge, including two sub-challenges (classification and automated segmentation), was posted. In this regard, Barker et al (6) demonstrated a novel method using local representative tiles, decided diagnostically by an elastic net classifier, for automated classification in brain cancer cases, which exhibited high accuracy for diagnosis as well as structural stability and a robustness for varying parameters, implying that it may be useful for automatic differentiation of the two subtypes of brain cancer (glioblastoma multiforme and lower grade glioma) (Table II). In another example, multiple-instance learning (MIL), first proposed by Dietterich et al in 1997, is a variation on supervised learning. In 2019, Campanella et al (7) presented the MIL-RNN models in the analysis of three datasets (including PCa, basal cell carcinoma of the skin and BC metastases to axillary lymph nodes datasets) of 44,732 WSIs from 15,187 patients, in which MIL-based residual neural network (ResNet)34 models were used to classify tiles. Then, semantically rich tile-level feature representations were generated, which integrated the information across the WSIs through the RNN models and achieved the final classification result. This system was shown to be able to train accurate classification models at an unprecedented scale, providing the cornerstone for the evolution of computational decision support systems in clinical practice (Fig. 1).

Notably, one of the biggest challenges in neuroanatomy [the automatic segmentation of neuronal structures presented in stacks of electron microscopy (EM) images] at ISBI 2012 was won by Ciresan et al (8) by employing a special type of deep ANN as a pixel classifier (Table I). However, there were two key disadvantages to the strategy raised by Ciresan et al, which were that it was slow and required a trade-off between localization accuracy and the use of context. Fortunately, a noteworthy finding was reported by Shelhamer et al (9) in 2017, where the fully convolutional networks (FCN) model could convert semantic-level images into pixel-level images using a convolution layer, instead of the full connection layer of the original segmentation network, which promoted the rapid development of image segmentation. For instance, Signaevsky et al (10) revealed that the FCN implemented in PyTorch was efficient and well suited for the practical application of WSIs derived from 22 autopsy brains from patients with tauopathies, which produced high precision and recall in naïve WSI semantic segmentation. Yi et al (11) also demonstrated that a finely-tuned FCN could be applied to analyze the microvessels of H&E stained histology images from patients with lung adenocarcinoma (LUAD) for prognostication. As a result of this FCN development, the U-Net model was presented in 2015 by Ronneberger et al (12) who, through modifying and extending the FCN, which supported automatic recognition and precise segmentation of characteristic lesions in EM stacks, beat the network presented by Ciresan et al (8) at ISBI 2015 (Table I). Furthermore, building on past efforts, Ronneberger et al developed an ImageJ plugin that enabled non-ML practitioners to solve problems with U-Net on either a local computer or a remote server/cloud service, which provided a generic DL-based software package for 2D and 3D cell detection and segmentation (12,13). A case in point, Lee et al (14) in 2020 used the modified U-Net to perform spatial analysis for the tumor microenvironment (TME) in WSIs from breast invasive carcinoma cases to replenish cancer classification and prediction, which demonstrated the effectiveness of the U-Net approach and the quantitative estimates derived from the spatial analysis. In addition, U-Net was proposed to aid the epidermis segmentation task on WSIs from malignant melanoma by Oskal et al (15), in which a superior performance of U-Net compared with existing techniques was observed.

In short, AI has been involved in the growth of DP, which has indicated that AI may bring a future of intelligent diagnostic robots.

In addition to the use of conventional pathology as the basis for DP, AI has also demonstrated unparalleled advantages and potential in MP, which is a sub-microscopic discipline of pathology. MP predicts risk, facilitates diagnosis and improves prognostication based on a complete understanding of the biological impact of specific molecular variants, mutations and dysregulations in diseases. MP techniques, which were rooted in fundamental molecular biology discoveries of the 1940s to 1980s (for example, PCR developed by Mullis in the 1980s), have moved from genomics to proteomics [such as next-generation sequencing, single cell RNA sequencing, fluorescence in situ hybridization, NMR spectroscopy and iTRAQ-MALDI-MS/MS], which has been driven by an increasing number of druggable targets and predictive biomarkers. Undeniably, AI fostered the development of MP, which was mainly encompassed by molecular profiles for risk prediction and molecular diagnostics, especially genomic diagnostics (16). For example, Woerl et al (17) formulated a novel ResNet-based approach (termed mibCNN) to predict the molecular subtype of muscle-invasive bladder cancer (MIBC) by analyzing two datasets of WSIs from H&E staining alone. The results demonstrated that DL could be trained to predict the significant molecular characteristics of MIBC from digital images derived from H&E slides only, potentially resulting in an improvement in managing this disease. Similarly, Schrammen et al (18) described a comparative performance with a state-of-the-art method (referred to as the slide-level assessment model), which simultaneously detected tumors and predicted the genetic alterations (including microsatellite instability/mismatch repair deficiency status and BRAF mutational status) of CRC from H&E-stained histopathological WSIs.

Numerous studies have focused on the risk stratification for cancer, especially for the globally leading cause of cancer deaths, LC. For instance, Choi and Na (19) proposed a novel DL-based risk stratification model to predict prognosis for patients with LUAD based on a gene co-expression network, which exhibited a significant association with patient prognosis independent of other clinicopathological features. Moreover, Choi et al (20) designed a robust Genomic Sequencing Classifier (GSC; a second-generation risk stratification algorithm) based on whole transcriptome RNA sequencing to validate the gene expression values predicting clinicopathological features of several LC cohorts. Compared with the Bronchial Genomic Classifier (the first generation), the GSC presented an optimized risk stratification across several independent LC cohorts, which aided physicians in capturing the optimal actionable information for the precision medicine of patients.

Growing evidence has shown that the TME, which is a complex physical and biochemical system comprising tumor cells, tumor stromal cells, vascular cells, immune cells and cytokines, plays an important role in tumor initiation, progression, metastasis and drug resistance. One of the most notable advances facilitated by AI has been in the understanding of the tumor-infiltrating lymphocytes (TILs) of the TME. Based on information regarding the TME from multiple studies in which immune content response data from The Cancer Genome Atlas (TCGA) were comprehensively analyzed, Saltz et al (21) framed a comprehensive methodology (termed Computational Staining) based on CNNs to map TILs and analyze the molecular and clinical correlation of TILs from H&E-stained WSIs containing 13 TCGA cancer types (Table II). In the present study, TIL map structural patterns with competitive performance were generated and it was revealed that the TIL patterns were associated with tumor and immune molecular features, cancer type and patient survival, which laid a new milestone for TME research (Fig. 3). Rakaee et al (22) described an attempt to associate an ML-based assessment of TILs from standard histological images with the outcomes of anti-PD-L1 monotherapy in patients with advanced non-small cell LC (NSCLC), which provided evidence that AI-based patient TIL assessment may enhance precision therapy.

As another key component of the TME, tumor-associated macrophages (TAMs) were studied by Bao et al (23) in triple-negative BC through an integrated analysis of single-cell and bulk RNA-seq. A DL approach based on the neural network, PyTorch, for the TAM-related gene signature was observed to display high accuracy in predicting the immunotherapy response (Fig. 3). Moreover, Cancian et al (24) demonstrated that CNN-based DeepLab-v3 could accurately recognize TAMs from the background, and separate different TAMs in marked contrast to the other two CNN-based models, which may satisfy the requirements of clinical practice for the characterization of TAM-related metrics in human colorectal liver metastases. In addition, Bian et al (25) designed a DL-based computational framework (termed ImmunoAIzer) to analyze the spatial distribution of immune cells and cancer cells in the TME and to detect gene mutations in colon cancer, which provided comprehensive information of cell distribution and tumor gene mutation status more efficiently and less expensively.

Moreover, tumor heterogeneity (including the contribution of the TME), an unneglectable factor affecting tumor evolution and therapy resistance, has also rapidly become a hot field of tumor phenotype research facilitated by AI. Undoubtedly, AI has made it easier to capture information from heterogeneous datasets to achieve accurate predictions and recognition patterns (26). For instance, since single-cell transcriptome sequencing technology was first reported by Tang et al (27) in 2009, and it has been widely applied in understanding molecular mechanisms and cellular properties of numerous biological processes, especially the tumor heterogeneity in various tumor types, such as CRC (28), ovarian cancer (29,30), head and neck cancer (31) and hepatocellular carcinoma (HCC) (32). In the light of this, Del Giudice et al (26) attempted to comprehensively review AI-based single-cell RNA-seq experiments in cancer from six network databases including the Genomic Data Commons Data Portal, ENCODE, Gene Expression Omnibus, Sequence Read Archive, Human Cell Atlas and Single Cell Portal, and sorted these experiments according to the tasks of AI models. These experiments, as well as the latest data, were categorized as defining cancer subtypes and cell clones (33–36), parsing tumor immune microenvironment (37–39), identifying new drug biomarkers (40), assessing and predicting disease recurrence and patient survival (41–45), detecting new putative actionable vulnerabilities (44,46) and predicting tumor immunoprofiling (47) (Table II) (26). More notably, the contributions of AI to the needs of cancer genomics for improving patient management were highlighted (26).

With the promise of continuing advances in the autonomous systems, explainable AI (XAI; the emerging generation of AI), which modified DL techniques to learn explainable features, has come to the fore in precision medicine, guaranteeing that predictions were more comprehensible and robust both in DP and MP. For example, Altini et al (48) tested the performance of NDG class activation mapping (NDG-CAM), a gradient-weighted CAM (one of the XAI applications) based method for nuclei detection in histopathology WSIs. NDG-CAM was observed to outperform in five datasets compared with other state-of-the-art methods. Moreover, Gimeno et al (49) developed a novel XAI method (multi-dimensional module optimization; MOM), which was applied to an acute myeloid leukemia (AML) cohort of 319 ex vivo tumor samples with 122 screened drugs and whole-exome sequencing. The predictive performance of MOM in AML was successfully validated in three different large-scale screening experiments with a therapeutic strategy based on the FLT3, CBFβ-MYH11 and NRAS status, which predicted patient response to quizartinib, trametinib, selumetinib and crizotinib (49). In addition, Meena and Hasija (50) described the outperformance of XGBoost ML models in identifying 23 significant genes of squamous cell cancer (SCC) of the skin, which may be the diagnostic and prognostic biomarkers of patients with SCC.

Taken together, AI has empowered MP by challenging itself, which mitigated various broken promises of precision medicine to improve the poor prognosis of different cancer types.

To date, integration of AI and the cancer pathology workflow has achieved some breakthroughs for precision medicine, these applications were summarized and classified into four groups according to their principal tasks in precision medicine: DA, CP, drug benefit (DB) and disease prevention (DPr) (Figs. 1B and 2). A review of the literature from Google scholar and PubMed was conducted using the following terms: ‘AI’ AND ‘pathology’ AND ‘cancer’ AND ‘diagnosis’, or ‘AI’ AND ‘pathology’ AND ‘cancer’ AND ‘therapy’, or ‘AI’ AND ‘pathology’ AND ‘cancer’ AND ‘prognosis’. or ‘AI’ AND ‘pathology’ AND ‘cancer’ AND ‘prevention’. In total, ~20,000 available texts (Fig. 1B) were retrieved (most published studies had an impact factor of ≥5) and these principal applications of AI-integrated cancer pathology were listed chronologically for the first time (to the best of our knowledge) in Table II.

While reducing the workload of oncologists, the vast majority of these studies were concerned with integration of AI methods into cancer pathology to improve DA, including identification, classification, detection and discrimination of cancer and other malignancies. This means that integration of AI and cancer pathology has been used primarily as an aid to cancer diagnosis, and a great deal of effort has been devoted to address the bottlenecks in improving DA, such as low speed and efficiency of scanning slides, image distortion, spatial distribution complexities, limited recognition or auto-recognition ability and inefficient analytical capabilities, as well as the shortfalls for process integration (i.e., tradeoffs between performance and new features), etc. (Table II). Thus, these reasonable efforts have served to enhance the roles of cancer pathology in precision medicine, as far as possible, but they are not sufficient to keep pace with the ever-growing needs of personalized treatment.

Although the body of literature in the field of integration of AI into cancer pathology for CP is relatively small, the prediction of cancer prognosis has gained more attention, which is mainly attributed to the growth of MP. It is well-known that a cancer prognosis typically involves multiple physicians from different specialties using different subsets of biomarkers and multiple clinical factors, including the age and general health of the patient, the location, type, grade and size of the cancer, as well as the tumor-node-metastasis staging of the tumor. There are three predictive foci concerned with cancer prognostication, including the prediction of cancer susceptibility (i.e., risk assessment), the prediction of cancer recurrence and the prediction of cancer survivability. Differentiating these cancer predictive foci from biomarkers and factors may pose a problem. In this regard, as a consequence of the outperformance of AI integrated into MP driven by the rapid development of cancer biomarkers, which cover a broad range of biochemical entities, such as nucleic acids, proteins, sugars, small metabolites and cytogenetic and cytokinetic parameters, as well as entire tumor cells found in bodily fluids, improving predictions of prognosis via these biomarkers were observed in 10 types of TCGA cancer (41), soft tissue sarcomas (44), BC (46), cervical cancer (CC) (61), bladder cancer (73), glioma (82,127), AML (129) and chronic myelogenous leukemia (130) (Table II). Therefore, AI has helped clinicians to accurately stratify risk in patients via the assessment of various cancer biomarkers.

Similarly, encouraged by the emergence of cancer biomarkers and corresponding technologies, there has been a gradual increase in the studies related to the integrations of AI and cancer pathology in predicting drug benefits by evaluating novel cancer biomarkers, including TIGIT (38), activation of BH3-only proteins (39), amino acid synthesis and interconversion (39), tumor human leukocyte antigen peptide (47), CTLA4 (66), oncotype DX and other gene expression (69), 21 genes (73), dysplasia (129), TOP1, PDIA4 and OGN (141) (Table II). Accordingly, these studies have provided compelling evidence for the feasibility of developing targeted therapies against these molecules for patients to obtain optimized medical treatment.

Regardless of the number of papers focusing on DPr in terms of integration of AI and cancer pathology is the least among the four main tasks, the influence of these studies on cancer precision medicine remains vast. As a part of precision medicine, cancer prevention, which is generalized as being ‘founded on describing the burden of cancer, identifying the causes, and evaluating and implementing preventive interventions’, means there is ‘a significant amount of work to be conducted’. Cancer prevention has achieved some success in reducing the global cancer incidence. The most convincing evidence is that tobacco control efforts and smoking cessation efforts as well as additional efforts, such as computed tomography for LC screening, since the 1960′s will continue to reduce global LC incidence (150). Due to the advances in understanding cancer biology, precision prevention approaches have been made feasible in various types of cancer with the assistance of AI, such as in CRC (80,128), BC (91), CC (96,107), liver cancer (128), LUAD (128) and renal cell cancer (131) (Table II). It is hypothesized that these attempts will reduce global cancer burden in the future.

The integration of AI into these practices has allowed technologies to handle large amounts of data related to cancer pathology, which has not only alleviated the workload pressure of oncologists but has also improved the accuracy of diagnosis and prediction of cancer prognosis and drug benefits, as well as promoted DPr for precision medicine.

At present, cancer remains a major public health problem worldwide, and cancer death was adversely affected by the coronavirus 2019 (COVID-19) global pandemic. Thus, there is a pressing need to develop novel and effective strategies for cancer precision medicine. Although growing evidence indicates that a combination of cancer pathology and AI can bring exciting changes to cancer precision medicine, a large number of technical, ethical and legal challenges still need to be appropriately addressed.

Firstly, there are four main limitations in terms of technology level in the development of AI in cancer pathology, including: The quality and standardization of pathological images and WSIs (especially the images from 2D to 3D, if possible, or even 4D), the optimization of data, the interpretability of data and the verification of algorithms. Therefore, more efforts than could possibly be reviewed in the present review have been devoted to address these problems. For example, Zheng et al (151) created a Graph-Transformer fusing a graph-based representation of a WSI, and developed a graph-based digital pathology visual transducer to predict disease grade, which outperformed compared with current state-of-the-art methods for WSI classification. It is hypothesized that these behaviors could benefit cancer pathology ultimately with faster and more accurate diagnoses, as well as faster screening. Moreover, Qiao et al (152) described the rationalized DL-structured illumination microscopy and lattice light sheet microscopy system to build a minimally invasive 2D/3D live cell imaging system for following intracellular dynamics and trajectory, which provided the possibility that a novel system may be created in the future by extending the aforementioned capabilities to monitor the therapeutic effect of the treatments to tumor cells in vivo safely. In addition, owing to the fact that the effectiveness of these systems was limited by the inability of the machine to explain its thoughts and actions to human users, XAI, a suite of ML techniques, was created to promote safety and clarity by showing how decisions are made in AI models, especially in critical tasks, such as drug screening with genetic events (49,50) and drug-drug interaction predictions (153).

Nevertheless, cancer is a highly complex disease involving a cascade of microscopic and macroscopic changes with mechanisms and interactions that are not yet fully understood. With the continuous discovery of novel cancer biomarkers and the innovation of technologies, cancer biomarkers that provide insights into the state and course of disease in the form of quantitative or qualitative measurements have made the fusion of multiple medical perspectives to guide patient management feasible, which AI has promoted in multi-modal patterns. Since Chaudhary et al (83) first employed a DL-based model to identify multi-omics features to predict HCC survival risk, AI-guided research linked to multi-omics has emerged. For instance, Feng et al (135) proposed and evaluated an AI radiopath-omics integrated model (termed RAPIDS) to predict pathological complete response (pCR) to neoadjuvant chemoradiotherapy in patients with locally advanced rectal cancer via pretreatment MRI and H&E stained WSIs. The results suggested that RAPIDS was able to predict pCR to neoadjuvant chemoradiotherapy, based on pretreatment radiopath-omics images, with high accuracy and robustness, which demonstrated a successful fusion of imagen-omics, path-omics and clinical data to predict patient prognosis (Fig. 4). Furthermore, Wang et al (137) built a nomogram model combining path-omics, radio-omics, protein-omics and clinical data to predict the postoperative outcome of patients with CRC and lung metastasis, which demonstrated an outperformance in risk prediction. Moreover, Vanguri et al (144) recently developed an ML-based multimodal model to predict immunotherapy response to PD-L1 blockade by capturing the information through a combination of imagen-omics, path-omics, genomics and clinical data from patients with advanced NSCLC. The model provided a quantitative rationale for using AI-guided multimodal integration to improve prediction of the drug benefits and outcomes in patients with NSCLC. Hence, synergies of cancer pathology with other medical tools could provide more information to the clinic to make an accurate and rapid decision in personalized treatments for patients. However, the well-known failed implementation of AI in oncology (the IBM Watson program at MD Anderson) is a reminder that multiple deployment challenges must be overcome prior to the integration of a pathology AI system into a clinical work environment (154,155). It is hypothesized that exploring the potential advantages of multimodal integration of path-omics, genomics, imagen-omics, protein-omics, epigenomics and other omics, as well as clinical data to make appropriate management decisions and improve patient outcomes may be the most challenging issue of cancer precision medicine in the future.

Notably, recent attempts to verify the capability of ChatGPT in effecting precision medicine have been reported since ChatGPT was released at the end of 2022, as it was reasonable to assume that using larger language models to enhance relationships through better communication would have a beneficial impact on patients (156,157). Sinha et al (158) assessed the capability of ChatGPT in solving higher-order reasoning regarding pathology and found that ChatGPT may offer meaningful responses but also current limitations.

Secondly, the vulnerability of ethics in the development of AI in cancer pathology cannot be ignored. With the introduction of WSI scanners, slides (as high-resolution digital images) can be captured, stored and transmitted electronically, and oncologists can annotate and label images as part of their reports. When the images and reports are stored in the cloud, the process allows third parties to share the information and associated data for research or other purposes. Moreover, oncologists have an obligation to keep identifiable patient information confidential (except with expressed consent), however, we consider that more could be done to ensure patient privacy and data security in this process. For instance, Chauhan and Gullapalli (159) posed three key foundational principles of ethical AI in the context of cancer pathology: Transparency, accountability and governance, to guide future practice. This viewpoint is appreciated, and it deserves popularizing in the clinic.

Thirdly, it has been recognized that the adoption of a new technology cannot be accomplished without some changes to the human and social aspects of the organization. The lack of global standardization has prevented the reusing of AI-integrated cancer pathology data to meet the broad range of patient safety and quality reporting requirements, which has been attributed to the following potential factors: i) The standardization for storage and retrieval of medical images; ii) the high standards and regulations that surround the cancer pathology laboratory; iii) the adoption of a global diagnostic system (i.e., as with the World Health Organisation); iv) the normalization of validation systems; v) the establishment of corresponding legal liabilities related to AI; and vi) the oncologist, etc. To facilitate the integration of any medical data from any data owner worldwide without violating privacy laws, Warnat-Herresthal et al (160) proposed a decentralized ML approach termed Swarm Learning (SL), which united edge computing, blockchain-based peer-to-peer networking and coordination, while maintaining confidentiality without the need for a central coordinator, thereby going beyond federated learning. They chose four cases of heterogeneous diseases (COVID-19, tuberculosis, leukemia and lung pathologies) with >16,400 blood transcriptomes derived from 127 clinical studies with non-uniform distributions of cases and controls and substantial study biases, as well as >95,000 chest X-ray images to validate the feasibility of using SL to develop disease classifiers using distributed data. The results demonstrated that SL classifiers outperform those developed at individual sites, which allows for further study in more cancer types (160). It is hoped that in addition to establishing new rules and laws to provide a comprehensive guarantee for the development of AI in oncology pathology, more options could be explored to integrate the medical data for precision medicine to the extent allowed by laws (Fig. 4).

In conclusion, the 21st century has witnessed tremendous breakthroughs in AI and cancer pathology, and most of these creations and innovations have improved and advanced understanding of these systems. As such, it is hoped to improve understanding of the concepts related to these systems and to help others realize the goals of precision medicine.