Exploration of Targeted Anti-tumor Therapy Exploration of Targeted Anti-tumor Therapy 2692-3114 Open Exploration 1002127 10.37349/etat.2023.00127 Review Artificial intelligence applications in pediatric oncology diagnosis https://orcid.org/0000-0002-4405-5711 Yang Yuhan 1 Zhang Yimao 1 Li Yuan 2 * Nardone Valerio Department of Pediatric Surgery, West China Hospital, Sichuan University, Chengdu 610041, Sichuan, China Laboratory of Digestive Surgery, State Key Laboratory of Biotherapy and Cancer Center, Department of Pediatric Surgery, West China Hospital, Sichuan University, Chengdu 610041, Sichuan, China University of Campania “L. Vanvitelli”, Italy Correspondence: Yuan Li, Laboratory of Digestive Surgery, State Key Laboratory of Biotherapy and Cancer Center, Department of Pediatric Surgery, West China Hospital, Sichuan University, Chengdu 610041, Sichuan, China. l13258389785@126.com 2023 28 02 2023 4 1 157 169 17 10 2022 30 12 2022 © The Author(s) 2023. 2023 This is an Open Access article licensed under a Creative Commons Attribution 4.0 International License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, sharing, adaptation, distribution and reproduction in any medium or format, for any purpose, even commercially, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.

Artificial intelligence (AI) algorithms have been applied in abundant medical tasks with high accuracy and efficiency. Physicians can improve their diagnostic efficiency with the assistance of AI techniques for improving the subsequent personalized treatment and surveillance. AI algorithms fundamentally capture data, identify underlying patterns, achieve preset endpoints, and provide decisions and predictions about real-world events with working principles of machine learning and deep learning. AI algorithms with sufficient graphic processing unit power have been demonstrated to provide timely diagnostic references based on preliminary training of large amounts of clinical and imaging data. The sample size issue is an inevitable challenge for pediatric oncology considering its low morbidity and individual heterogeneity. However, this problem may be solved in the near future considering the exponential advancements of AI algorithms technically to decrease the dependence of AI operation on the amount of data sets and the efficiency of computing power. For instance, it could be a feasible solution by shifting convolutional neural networks (CNNs) from adults and sharing CNN algorithms across multiple institutions besides original data. The present review provides important insights into emerging AI applications for the diagnosis of pediatric oncology by systematically overviewing of up-to-date literature.

 Pediatric oncology artificial intelligence cancer diagnosis machine learning deep learning Introduction

The exponentially growing knowledge and techniques have explored innovative perspectives for multi-layered diagnoses in pediatric oncology that the expectation of patients and their families have been developed for their specific situation to receive optimized care instantly and comprehensively. Artificial intelligence (AI) strategies can tackle enormous amounts of original data in a short time to solve complex tasks with high accuracy [1, 2]. Physicians can improve their diagnostic efficiency with the assistance of AI techniques for improving the subsequent personalized treatment and surveillance.

AI has developed into various computer-assisted theories and is mainly implemented with working principles of machine learning (ML) and deep learning (DL). AI algorithms fundamentally capture data, identify underlying patterns, achieve preset endpoints, and provide decisions and predictions about real-world events. ML, as a main subset of AI, indicates a different flowchart compared with traditional hard-coded software programs which apply algorithms to construct predictive models dynamically by training large amounts of historical data. DL, as a growing aspect of AI, represents benefits in learnable weights and high efficiency with minimal pre-processing based on the structure of convolutional neural networks (CNNs) with multiple inter-connected layers. DL-CNNs composed of multiple stacked CNN layers have advantages in accurate, faster, vendor-independent processing compared with ML algorithms applied previously [3]. Notably, AI methods have been demonstrated to offer effective assistance for clinical management, including cancer segmentation, susceptibility, and classification, as essential fundamental for early diagnosis and prognosis management in cancer research [4, 5].

The application of AI not only makes full use of the various aspects of clinical diversity but also helps to address the current lack of objectivity and universality in expert systems [6]. The application of AI can help hospitals train junior physicians in clinical diagnosis and decision-making. A growing number of research papers are reporting the impressive diagnostic performance of computer systems built using ML [7]. DL techniques, in particular, are transforming their ability to interpret imaging data [8, 9]. These results may improve sensitivity and ensure fewer false positives than radiologists. However, DL techniques also run the risk of overfitting the training data, resulting in a brittle degraded performance in certain settings [10]. Thus, AI often involves a tradeoff between accuracy and intelligibility.

Despite the current AI advancements, the sustainable development of health AI tools relies on the availability of large datasets with strict quality control [11]. Several biomedical imaging repositories have been created to date [12, 13], such as The Cancer Imaging Archive (TCIA) one of the most renowned repositories focusing on cancer imaging [14]. Nowadays, there have been specific data repositories for pediatric cancer, for example, the PRIMAGE project, as an open cloud-based platform based on European populations, involves high-quality multidimensional anonymized datasets (imaging, clinical, molecular, and genetics) for the training and validation of ML and multiscale algorithms [15]. Albeit of huge potential, the vast majority of these repositories have been created as stand-alone entities, being currently not in a position to become interoperable with similar existing initiatives. As such, the creation of a fully findable, accessible, interoperable, reusable (FAIR) repository based on multiple populations is still warranted for AI analysis [16].

The prosperous development of AI techniques promotes numerous potential applications in pediatric oncology with two main bottlenecks for successful utilities, including the need for large data entry and a strong graphic processing unit (GPU) with appropriate computer and memory power. AI algorithms provide timely references based on large amounts of clinical and imaging data and sufficient GPU power [17]. Meanwhile, DL-CNNs can learn from medical literature automatically to capture innovative and feasible ideas for literature reviews, which can assist in accurate diagnosis at an early stage and optimal treatment selection for individuals [18]. Nowadays, it’s necessary to feed AI algorithms with representative large-sample data sets, hundreds or thousands for typical cases where sparse and/or unqualified data induces unsatisfactory performance and unreliable outcomes [18]. Currently, neither ML nor DL can guarantee accuracy and consistency in specific circumstances without similarity compared with historical training data. This is an inevitable challenge especially for pediatric oncology because of the low morbidity and individual heterogeneity. However, this problem may be solved in the near future considering the exponential advancements of AI algorithms technically to decrease the dependence of AI operation on the amount of data sets and the efficiency of computing power. For instance, it could be a feasible solution by shifting CNNs from adults and sharing CNN algorithms across multiple institutions besides original data. The present review provides important insights into emerging AI applications for the diagnosis of pediatric oncology by systematically overviewing updated literature.

 Current AI applications in the diagnosis of pediatric oncology

Clinicians make clinical diagnoses depending on their professional knowledge and clinical experience through signs, symptoms, laboratory, and imaging examinations. It seems hard to guarantee diagnostic accuracy and consistency by dealing with abundant and multidimensional data from the human brain. The AI algorithms have advantages in learning and training vast amounts of data and integrating it into a certain outcome in a very short time, which allows efficiency and effectiveness of computer-aided diagnosis (CAD) in clinical practices. DL algorithms have recently made significant progress in extracting and processing information from medical images, which have been applied in various medical tasks extensively not only in radiology and pathology with satisfactory performance comparable to or even superior to that of human experts. Notably, DL algorithms could identify underlying information from medical images associated with tumor diagnosis [19]. The AI applications in the diagnosis of pediatric oncology are summarized in specific cancer types using ML and/or DL methods.

Non-solid tumor diagnosis

A performance comparison of diagnosis in pediatric hematological malignancies using AI strategies is shown in Table 1. The results of cluster and discriminant analyses for various types of pediatric acute leukemia revealed that a combination of DL analysis and microscopic blood images facilitated the classification of acute leukemia and outperformed expert hematologists with an accuracy of more than 98% [20, 21]. The utility of AI in the automatic analysis of microscopy images represented diagnostic accuracy of around 95% in acute promyelocytic leukemia [22], acute lymphoblastic leukemia (ALL) [23, 24], and leukemic B-lymphoblast [25], which was optimized by a hybrid model using a genetic algorithm and a residual CNN reaching an accuracy of 98.46% [26]. The DL analysis was also applied in the classification of ALL, acute myeloid leukemia (AML), and chronic myeloid leukemia (CML) using bone marrow cell microscopy images [27]. Otherwise, an easy-to-interpret transcriptome-wide biomarker was developed for accurate ALL subtyping by elucidating diagnostic associations between messenger RNA (mRNA) sequencing profiles and ALL lesions [28]. The ML-based strategies on DNA methylation showed advantages in the differentiation of leukemia blood from normal blood [29, 30].

Current AI applications in the diagnosis of childhood non-solid tumor

AI method Medical field Task Tumor Result References
CNN/GAN Pathology Detecting ALL and AML using a deep learner classifier using microscopic blood images ALL and AML ACC: 98%–98.67% [20, 21]
CNN and GAN Pathology/Genomics Constructing a hybrid model using a genetic algorithm and a residual CNN to predict ALL using microscopy images ALL ACC: 98.46% [26]
SVM Pathology Building a model to classify acute leukemias using flow cytometry Acute promyelocytic leukemia ACC: 94.2%; AUC: 99.5 [22]
ANN/FFNN/SVM Pathology Proposing a ML-based model for ALL categorization using microscopic blood images ALL ACC: 98.1–100% [23, 24]
CNN Pathology Building an aggregated DL model for leukemic B-lymphoblast classification Leukemic B-lymphoblast ACC: 96.58% [25]
CNN Pathology Using bone marrow cell microscopy images for the classification of AML, ALL, and CML AML, ALL, and CML ACC: 90–99% [27]
RF Others-mRNA sequencing Developing transcriptome-wide biomarkers for ALL subtyping ALL ACC: 97–100% [28]
ANN Others-DNA methylation Identifying reliable cancer-associated methylation signals in gene regions from leukemia patients Leukemia ACC: 93.8% [29]
Nearest shrunken centroids Others-DNA methylation Investigating the utility of CpG methylation status to differentiate blood from patients with ALL and AML from normal blood ALL and AML AUC: 99.98 [30]

GAN: generative adversarial network; SVM: support vector machine; ANN: artificial neural network; FFNN: feed forward neural network; ACC: accuracy; AUC: area under the curve; RF: random forest; CpG: cytosine-guanine

 Solid tumor diagnosis Intracranial tumor diagnosis

A performance comparison of diagnosis in pediatric intracranial tumors using AI strategies is shown in Table 2. Recently, the use of an ML-based classification model has been shown to improve the diagnostic accuracy of childhood intracranial tumors, especially for posterior fossa tumors by analyzing multiple magnetic resonance imaging (MRI) sequences that the DL-CNN model showed improved accuracy with similar sensitivity and improved specificity in terms of discriminate several hard-to-differentiate brain tumor subtypes compared to radiologists’ recognition [2]. The improved performance of these AI-based models warranted further verification in prospective studies and even randomized clinical trials for broad clinical practice. For different tumor subgroups coupled with overlapping neuroimaging features, the use of DL-CNNs could identify these atypical cases with high discriminative performance (AUC: 0.81–0.98), which provided low-cost and high-efficiency decision-making support for reference of diagnostic brain biopsy or maximal tumor resection [31–40]. Moreover, a novel gene-derived algorithm improved the performance of AI-based models with small-scale data by identifying optimal architectures using feature embeddings from state-of-the-art image classification networks that this approach might offer an available solution for the small-sample task in pediatric brain tumors [41].

Current AI applications in the diagnosis of childhood intracranial tumor

AI method Medical field Task Tumor Result References
SVM Radiology Classifying pediatric posterior fossa tumors Pediatric posterior fossa tumors ACC: 75–85% [32]
LR/XGB/LASSO Radiology Distinguishing pediatric supratentorial tumors, high-grade gliomas, and ependymomas Supratentorial embryonal tumors, high-grade gliomas, and ependymomas ACC: 81–91%; AUC: 0.82–0.98 [33]
CNN Radiology Establishing a pre-trained ResNet18 with transfer learning to identify germinomas of the basal ganglia Gliomas and germinomas AUC: 0.88 [34]
CNN Radiology Training on T1-weighted gadolinium-enhanced MRI scans of glioblastomas, atypical primary central nervous system lymphomas, and solitary brain metastasis Glioblastomas, atypical primary central nervous system lymphomas, and solitary brain metastasis AUC: 0.81–0.98 [31]
CNN/RF/DT/KNN/SVM Radiology Classifying primary central nervous system lymphoma and glioma types Primary central nervous system lymphoma and glioma ACC: 84%; AUC: 0.839 [35]
LR/ANN Radiology Developing a sequential ML classifier to distinguish medulloblastoma from ependymoma Medulloblastoma and ependymoma ACC: 94–95.5% [36]
SVM/LR/KNN/RF/XGB/ANN Radiology Distinguishing atypical teratoid/rhabdoid tumors and medulloblastomas by MR imaging-based radiomic phenotypes Atypical teratoid/rhabdoid tumors and medulloblastomas ACC: 81%; AUC: 0.86 [37]
CNN Radiology Characterizing and classifying multiple tumor histologic features in pediatric high-grade brain tumors employing diffusion basis spectrum imaging Pediatric high-grade brain tumors AUC: 0.950–0.991 [38]
GBDT Radiology Applying multiparametric MRI to differentiate pilocytic astrocytoma from cystic oligodendrogliomas Pilocytic astrocytoma and cystic oligodendrogliomas AUC: 0.99 [39]
CNN Radiology Developing an MR imaging-based DL model for posterior fossa tumor detection and tumor pathology classification Diffuse midline glioma of the pons, medulloblastoma, pilocytic astrocytoma, and ependymoma ACC: 92%; AUC: 0.99 [40]
CNN Radiology Identifying the pediatric brain tumor, adamantinomatous craniopharyngioma Adamantinomatous craniopharyngioma ACC: 83.3–87.8% [41]
CNN Pathology Proposing a time-efficient and reliable CAD for the automatic diagnosis of pediatric medulloblastoma and its subtypes from histopathological images Medulloblastoma ACC: 90–100% [42, 43]
SVM Others-blood markers Differentiating malignant and benign pediatric brain tumors using blood markers Pediatric brain tumors ACC: 71.6% [45]
LR Others-Raman spectroscopy Investigating the potential for Raman spectroscopy to accurately diagnose pediatric brain tumors intraoperatively Pediatric brain tumors AUC: 0.91–0.94 [46]
RF Others-DNA methylation profiles Describing a fast and cost-efficient workflow for intraoperative classification of brain tumors based on DNA methylation profiles generated by low coverage nanopore sequencing and ML algorithms Pediatric brain tumors ACC: 89% [47]
SVM Others-proteomics of cerebrospinal fluid Distinguishing among brain tumor versus non-tumor/hemorrhagic conditions and differentiating two large classes of brain tumors Pediatric brain tumors AUC: 0.97–1 [48]
LASSO Others-lncRNAs Developing an RF-based ML algorithm identifying a lncRNA-based diagnostic signature Medulloblastoma AUC: 0.974–1 [44]

LR: logistic regression; XGB: extreme gradient boosting; LASSO: least absolute shrinkage and selection operator; ResNet18: residual neural network with 18-layer by using more 5-layer blocks; DT: decision tree; KNN: k-nearest neighbour; GBDT: gradient boosting decision tree; lncRNAs: long non-coding RNAs

Attallah [42, 43] proposed two time-efficient and reliable CAD systems called MB-AI-His and CoMB-Deep for the automatic diagnosis of pediatric medulloblastoma and its subtypes based on histopathological whole-slide images. These systems combined the benefits of DL techniques and textural analysis feature extraction methods through a cascaded manner to achieve diagnostic and time efficiency simultaneously. The subtypes of pediatric medulloblastoma could also be accurately identified by an RF-based ML algorithm based on 11 lncRNA variables [44]. Otherwise, some innovative techniques were applied to diagnose pediatric brain tumors using AI techniques intraoperatively. Khayat Kashani et al. [45] used inflammatory indicators on peripheral blood tests to classify pediatric benign and malignant brain tumors. Jabarkheel et al. [46] explored the intraoperative diagnostic potential of Raman spectroscopy using an ML classifier in pediatric brain tumors. This ML-based method differentiated the normal brain from neoplastic tissue in a non-invasive manner and efficient diagnostic performance (AUC > 0.90) compared with microscopic visualization and intraoperative navigation. Djirackor et al. [47] applied DNA methylation nanopore sequencing and ML algorithms for the intraoperative classification of brain tumors. This approach performed correct diagnosis in all six cases with a median operating time of 97 min, which assisted decision-making for surgeons within a timeframe. Bruschi et al. [48] used waste cerebrospinal fluid from extraventricular drainage and used proteome analysis to distinguish patients with brain tumors versus non-tumor/hemorrhagic and classify subtypes of brain tumors.

 Extracranial tumor diagnosis

AI techniques have been applied in the diagnosis of pediatric extracranial tumors, mainly including soft-tissue and bone tumors (Table 3). For bone tumors, AI-based algorithms were used to classify benign and malignant lesions based on multiple radiological images, such as X-ray and MRI [49]. He et al. [50] and Pan et al. [51] proposed DL-CNN models with similar accuracy compared to subspecialists and better performance than junior radiologists. The same strategy was applied in pediatric soft-tissue sarcoma a ML-based model yielded great discriminative performance for differentiation between malignant and benign soft-tissue masses with an accuracy of 90.5%, sensitivity of 100%, and specificity of 80.6% [52] as well as other types of sarcomas versus benign diseases [53, 54]. The AI-based techniques on radiological images have been only applied in extremity tumors and the diagnostic potential of AI-based strategies needs further exploration in other tumors, such as abdominal and chest tumors.

Current AI applications in the diagnosis of childhood extracranial tumor

AI method Medical field Task Tumor Result References
CNN Radiology Developing an AI algorithm to distinguish osteomyelitis from Ewing sarcoma Ewing sarcoma and osteomyelitis ACC: 86.7–94.4% [53]
CNN/SVM Radiology Constructing image-based models to identify well-differentiated liposarcoma and lipoma Well-differentiated liposarcomas and lipomas ACC: 86.84%; AUC: 0.942 [54]
CNN/RF Radiology Developing a DL/ML model to classify primary bone tumors from preoperative radiographs and compare performance with radiologists Malignant and benign bone tumors AUC: 0.79–0.97 [49–51]
SVM/GLM/RF Radiology Constructing a radiomics-based machine method for differentiation between malignant and benign soft-tissue masses Malignant and benign soft-tissue masses AUC: 0.88–0.96; ACC: 80.8–90.5% [52]
CNN Pathology Building CNNs for rhabdomyosarcoma histology subtype classification Rhabdomyosarcoma AUC: 0.92–0.94 [55]
CNN Pathology Developing a DL CNN-based differential diagnosis system on soft-tissue sarcoma subtypes based on whole histopathology tissue slides Soft-tissue sarcoma AUC: 0.889 [56]
LDA Pathology Identifying proteomic differences, which would more reliably differentiate between benign and malignant melanocytic lesions Benign nevi and melanomas SEN: 98.76%; SPE: 99.65% [59]
CNN Others-dermatological photos Establishing an AI algorithm to diagnose infantile hemangiomas based on clinical images Infantile hemangiomas ACC: 91.7% [61]
LR Others-umbilical cord blood sera Exploring prediction biomarkers for infantile hemangiomas using noninvasive umbilical cord blood Infantile hemangiomas AUC: 0.756–0.943 [62]
CNN Others-dermoscopic examination Developing a DCNN model to support dermatologists in the classification and management of atypical melanocytic skin lesions Early melanomas and atypical nevi AUC: 0.903 [60]
SVM Others-cell-free DNA Providing a comprehensive analysis of circulating tumor DNA beyond recurrent genetic aberrations for early diagnosis Ewing sarcoma and other pediatric sarcomas SEN: 73%; SPE: 100% [57]
SVM Others-electronic colorimeters Determining the diagnostic utility of widely available colorimetric technology when differentiating port-wine birthmarks from infantile hemangiomas in photographs of infants less than 3 months old Port-wine birthmarks and infantile hemangiomas ACC: 90% [63]
DT & RF Others-array-generated DNA methylation data Classifying soft tissue and bone tumors using an ML classifier algorithm based on array-generated DNA methylation data Soft tissue and bone tumors AUC: 0.999 [58]

GLM: general linear model; LDA: linear discriminant analysis; DCNN: deep CNN

Zhang et al. [55] and Frankel et al. [56] developed DL-CNN differential diagnosis system with an AUC of 0.889 for pre-pathologist screening and quantifying diagnosis likelihood of trained soft-tissue sarcoma subtypes based on whole histopathology tissue slides. Considering the limited worldwide availability of sarcoma pathology expertise, this AI-based approach suggested assistance for local pathologists to quickly narrow the differential diagnosis of the sarcoma subtype in children, adolescents, and young adults. The cell-free DNA of Ewing sarcoma and other pediatric sarcomas for liquid biopsy could achieve sensitive detection and classification in peripheral blood independent of any genetic alterations [57]. The potential of array-generated DNA methylation data in early-stage diagnosis and classification was explored based on an ML-based classifier algorithm [58]. Otherwise, the CAD strategy was applied in the detection and diagnosis of childhood dermatological tumors using histopathological and dermoscopic data. Lazova et al. [59] performed ML-based histopathology-guided mass spectrometry profiling analysis on formalin-fixed, paraffin-embedded tissue samples to identify the proteomic level and achieve a sensitivity of 98.76% and specificity of 99.65% in determining benign nevi and malignant melanomas. Tognetti et al. [60] applied dermoscopic images to develop a DL-CNN model in the classification between atypical nevi from early melanomas, which achieved adequate accuracy (AUC: 90.3, sensitivity: 86.5%, and specificity: 73.6%) and eliminated the influence from dermatologists’ experience. Several AI-based algorithms on novel resources also achieved accurate diagnostic performance for skin neoplasms including photos of skin neoplasms [61], umbilical cord blood sera [62], and electronic colorimeters [63].

 AI applications in the identification of diagnostic signature

AI-based principles have been used for the detection and segmentation of pediatric malignant tumors. For example, Wu et al. [64] used a residual fusion network to detect osteosarcomas on MRI scans; Peng et al. [65] used a CNN for automated pediatric brain tumor detection and segmentation on MRI scans with automatic two-dimensional (2D) and volumetric size measurement of tumors; Strijbis et al. [66] used a CNN for automated eye and tumor segmentation on MRI in retinoblastoma patients; and Bouget et al. [67] used three-dimensional neural network architectures to automatically detect meningioma on MRI scans.

The AI algorithms have represented strengths in detecting tumor patterns by identifying underlying genetic and molecular characteristics associated with specific macroscopic tumor features based on medical images and/or high-throughput data. Zhao et al. [32] reported that ML-based radiomics algorithms could predict H3 K27M amplification status in children with midline glioma with significantly greater accuracy ranging from 0.788 to 0.867 than prediction by chance. Giwa et al. [68] applied ML-based algorithms in predicting N-myc status and survival risk using CpG methylation in children with neuroblastoma.

 Challenges and future outlook

The application of AI technology faces some important challenges that must be resolved to ensure its use in pediatric cancer diagnosis [69]. For example, medical imaging data cannot be used as input data directly. It is crucial to extract features from the imaging data and process them. Development and popularization of technology, in addition, the weights coefficient in CNN models are tested, calculated, and the confidence interval is reasonable, so medical interpretation needs further research [70]. While the importance of AI to this field is recognized, the joint efforts of computer experts and medical experts toward ensuring interdisciplinary personnel training and collaboration are crucial. Only then can the potential of this technology be put to a practical and economic application by medical staff [71]. The possibility is that the “black box” of ML/CNN applications will reduce physician skills and soon transform some sectors of healthcare in ways that may appear to be practical and economic but with unintended negative consequences. Another crucial issue with regard to the future of AI in medicine involves privacy and data security assurances [72]. While recent years have witnessed much enthusiasm about the potential of “big data” and ML-based solutions, to date, only a few examples exist to illustrate the impact of AI on current clinical practice [73, 74]. The stimulating debate that whether AI is “smarter” than human practitioners is largely irrelevant, and we will consistently improve our collective health by using every information and data resource [10].

 Conclusions

AI techniques have revolutionized the diagnostic field of oncology. Although AI approaches have been widely implemented in adult tumors, specialized applications of AI algorithms in childhood cancer are still limited probably attributed to the insufficient amounts of available data sets. There are limited opportunities to transfer well-trained CNN architectures built on adults into pediatric oncology few CNNs are directly generalizable from adults to children. Therefore, it’s warranted urgently to develop dedicated AI algorithms applied in pediatric oncology. Although the data sets of pediatric oncology are not large enough to perform standardized DL analysis in medical imaging [75], the integration of detailed segmentation and standardized augmentation techniques [76] is expected to achieve satisfactory performance proven by recent literature on brain tumors [77]. Combined with a considerable amount of pediatric cancer lesions, DL-CNNs yield reasonable and promising performance in the field of pediatric oncology considering its nature of pixel-level classification. Based on currently collected data, the radiological images of brain cancers have great potential for AI implementation, which might be validated first for the assistance of clinical decisions in the near future. Further improvement of pediatric oncology diagnosis requires sharing of source data and algorithms among multiple institutions for standardizations and cross-validation to benefit subsequent treatment and surveillance maximally.

 Abbreviations AI:

artificial intelligence

 ALL:

acute lymphoblastic leukemia

 AML:

acute myeloid leukemia

 AUC:

area under the curve

 CAD:

computer-aided diagnosis

 CML:

chronic myeloid leukemia

 CNNs:

convolutional neural networks

 DL:

deep learning

 lncRNAs:

long non-coding RNAs

 ML:

machine learning

 MRI:

magnetic resonance imaging

 RF:

random forest

 Declarations Author contributions

YY, YZ, and YL: Conceptualization, Writing—original draft, Writing—review & editing. YL: Validation, Writing—review & editing, Supervision.

 Conflicts of interest

The authors declare that they have no conflicts of interest.

 Ethical approval

Not applicable.

 Consent to participate

Not applicable.

 Consent for publication

Not applicable.

 Availability of data and materials

The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request. Requests for access to these datasets should be directed to [Yuan Li, l13258389785@126.com].

 Funding

Not applicable.

 Copyright

© The Author(s) 2023.

 References CallawayECastelvecchiDCyranoskiDGibneyELedfordHLeeJJ 2017 in news: the science events that shaped the year. Nature. 2017;552:304–7. 10.1038/d41586-017-08493-x 29293246 HuangJShlobinNALamSKDeCuypereM. Artificial intelligence applications in pediatric brain tumor imaging: a systematic review. World Neurosurg. 2022;157:99–105. 10.1016/j.wneu.2021.10.068 34648981 BanerjeeICrawleyABhethanabotlaMDaldrup-LinkHERubinDL. Transfer learning on fused multiparametric MR images for classifying histopathological subtypes of rhabdomyosarcoma. Comput Med Imaging Graph. 2018;65:167–75. 10.1016/j.compmedimag.2017.05.002 28515009 KourouKExarchosTPExarchosKPKaramouzisMVFotiadisDI. Machine learning applications in cancer prognosis and prediction. Comput Struct Biotechnol J. 2015;13:8–17. 10.1016/j.csbj.2014.11.005 25750696 PMC4348437 TanAHuangHZhangPLiS. Network-based cancer precision medicine: a new emerging paradigm. Cancer Lett. 2019;458:39–45. 10.1016/j.canlet.2019.05.015 31125640 PedersenMHHoodBLEhmsenSBeckHCConradsTPBakM CYPOR is a novel and independent prognostic biomarker of recurrence-free survival in triple-negative breast cancer patients. Int J Cancer. 2019;144:631–40. 10.1002/ijc.31798 30110125 FerroniPZanzottoFMRiondinoSScarpatoNGuadagniFRoselliM. Breast cancer prognosis using a machine learning approach. Cancers (Basel). 2019;11:328. 10.3390/cancers11030328 30866535 PMC6468737 ChartrandGChengPMVorontsovEDeozdzalMTurcotteSPalCJ Deep learning: a primer for radiologists. Radiographics. 2017;37:2113–31. 10.1148/rg.2017170077 29131760 LanKWangDTFongSLiuLSWongKKLDeyN. A survey of data mining and deep learning in bioinformatics. J Med Syst. 2018;42:139. 10.1007/s10916-018-1003-9 29956014 ChenJHAschSM. Machine learning and prediction in medicine—beyond the peak of inflated expectations. N Engl J Med. 2017;376:2507–9. 10.1056/NEJMp1702071 28657867 PMC5953825 LambinPRios-VelazquezELeijenaarRCarvalhoSvan StiphoutRGGrantonP Radiomics: extracting more information from medical images using advanced feature analysis. Eur J Cancer. 2012;48:441–6. 10.1016/j.ejca.2011.11.036 22257792 PMC4533986 SudlowCGallacherJAllenNBeralVBurtonPDaneshJ UK biobank: an open access resource for identifying the causes of a wide range of complex diseases of middle and old age. PLoS Med. 2015;12:e1001779. 10.1371/journal.pmed.1001779 25826379 PMC4380465 BambergFKauczorHUWeckbachSSchlettCLForstingMLassSC Whole-body MR imaging in the German national cohort: rationale, design, and technical background. Radiology. 2015;277:206–20. 10.1148/radiol.2015142272 25989618 ClarkKVendtBSmithKFreymannJKirbyJKoppelP The cancer imaging archive (TCIA): maintaining and operating a public information repository. J Digit Imaging. 2013;26:1045–57. 10.1007/s10278-013-9622-7 23884657 PMC3824915 Martí-BonmatíLAlberich-BayarriÁLadensteinRBlanquerISegrellesJDKeimD PRIMAGE project: predictive in silico multiscale analytics to support childhood cancer personalised evaluation empowered by imaging biomarkers. Eur Radiol Exp. 2020;4:22. 10.1186/s41747-020-00150-9 32246291 PMC7125275 GabelloniMFaggioniLBorgheresiRRestanteGShortredeJTumminelloL Bridging gaps between images and data: a systematic update on imaging biobanks. Eur Radiol. 2022;32:3173–86. 10.1007/s00330-021-08431-6 35001159 EricksonBJKorfiatisPAkkusZKlineTL. Machine learning for medical imaging. Radiographics. 2017;37:505–15. 10.1148/rg.2017160130 28212054 PMC5375621 KangJRancatiTLeeSOhJHKernsSLScottJG Machine learning and radiogenomics: lessons learned and future directions. Front Oncol. 2018;8:228. 10.3389/fonc.2018.00228 29977864 PMC6021505 LevineABSchlosserCGrewalJCoopeRJonesSJMYipS. Rise of the machines: advances in deep learning for cancer diagnosis. Trends Cancer. 2019;5:157–69. 10.1016/j.trecan.2019.02.002 30898263 ShawlyTAlsheikhyAA. Biomedical diagnosis of leukemia using a deep learner classifier. Comput Intell Neurosci. 2022;2022:1568375. 10.1155/2022/1568375 36072723 PMC9444372 KararMEAlotaibiBAlotaibiM. Intelligent medical IoT-enabled automated microscopic image diagnosis of acute blood cancers. Sensors (Basel). 2022;22:2348. 10.3390/s22062348 35336523 PMC8949784 MonaghanSALiJLLiuYCKoMYBoyiadzisMChangTY A machine learning approach to the classification of acute leukemias and distinction from nonneoplastic cytopenias using flow cytometry data. Am J Clin Pathol. 2022;157:546–53. 10.1093/ajcp/aqab148 34643210 AbunadiISenanEM. Multi-method diagnosis of blood microscopic sample for early detection of acute lymphoblastic leukemia based on deep learning and hybrid techniques. Sensors (Basel). 2022;22:1629. 10.3390/s22041629 35214531 PMC8876170 JawaharMHSLJAGandomiAH. ALNett: a cluster layer deep convolutional neural network for acute lymphoblastic leukemia classification. Comput Biol Med. 2022;148:105894. 10.1016/j.compbiomed.2022.105894 35940163 KasaniPHParkSWJangJW. An aggregated-based deep learning method for leukemic B-lymphoblast classification. Diagnostics (Basel). 2020;10:1064. 10.3390/diagnostics10121064 33302591 PMC7763941 RodriguesLFBackesARTravençoloBANde OliveiraGMB. Optimizing a deep residual neural network with genetic algorithm for acute lymphoblastic leukemia classification. J Digit Imaging. 2022;35:623–7. 10.1007/s10278-022-00600-3 35199257 HuangFGuangPLiFLiuXZhangWHuangW. AML, ALL, and CML classification and diagnosis based on bone marrow cell morphology combined with convolutional neural network: a STARD compliant diagnosis research. Medicine (Baltimore). 2020;99:e23154. 10.1097/MD.0000000000023154 33157999 PMC7647529 MäkinenVPRehnJBreenJYeungDWhiteDL. Multi-cohort transcriptomic subtyping of B-cell acute lymphoblastic leukemia. Int J Mol Sci. 2022;23:4574. 10.3390/ijms23094574 35562965 PMC9099612 SanchezRMackenzieSA. Integrative network analysis of differentially methylated and expressed genes for biomarker identification in leukemia. Sci Rep. 2020;10:2123. 10.1038/s41598-020-58123-2 32034170 PMC7005804 JiangHOuZHeYYuMWuSLiG DNA methylation markers in the diagnosis and prognosis of common leukemias. Signal Transduct Target Ther. 2020;5:3. 10.1038/s41392-019-0090-5 32296024 PMC6959291 TariciottiLCaccavellaVMFioreGSchisanoLCarrabbaGBorsaS A deep learning model for preoperative differentiation of glioblastoma, brain metastasis and primary central nervous system lymphoma: a pilot study. Front Oncol. 2022;12:816638. 10.3389/fonc.2022.816638 35280801 PMC8907851 ZhaoDGristJTRoseHELDaviesNPWilsonMMacPhersonL Metabolite selection for machine learning in childhood brain tumour classification. NMR Biomed. 2022;35:e4673. 10.1002/nbm.4673 35088473 ZhangMTamLWrightJMohammadzadehMHanMChennE Radiomics can distinguish pediatric supratentorial embryonal tumors, high-grade gliomas, and ependymomas. AJNR Am J Neuroradiol. 2022;43:603–10. 10.3174/ajnr.A7481 35361575 YeNYangQChenZTengZLiuPLiuX Classification of gliomas and germinomas of the basal ganglia by transfer learning. Front Oncol. 2022;12:844197. 10.3389/fonc.2022.844197 35311111 PMC8928458 LuGZhangYWangWMiaoLMouW. Machine learning and deep learning CT-based models for predicting the primary central nervous system lymphoma and glioma types: a multicenter retrospective study. Front Neurol. 2022;13:905227. 10.3389/fneur.2022.905227 36110392 PMC9469735 ZhangMWongSWWrightJNToescuSMohammadazdehMHanM Machine assist for pediatric posterior fossa tumor diagnosis: a multinational study. Neurosurgery. 2021;89:892–900. 10.1093/neuros/nyab311 34392363 PMC8764569 ZhangMWongSWLummusSHanARAhmadianSSProloLM Radiomic phenotypes distinguish atypical teratoid/rhabdoid tumors from medulloblastoma. AJNR Am J Neuroradiol. 2021;42:1702–8. 10.3174/ajnr.A7200 34266866 PMC8423034 YeZSrinivasaKMeyerASunPLinJVioxJD Diffusion histology imaging differentiates distinct pediatric brain tumor histology. Sci Rep. 2021;11:4749. 10.1038/s41598-021-84252-3 33637807 PMC7910493 ZhaoYLuYLiXZhengYYinB. The evaluation of radiomic models in distinguishing pilocytic astrocytoma from cystic oligodendroglioma with multiparametric MRI. J Comput Assist Tomogr. 2020;44:969–76. 10.1097/RCT.0000000000001088 32976261 QuonJLBalaWChenLCWrightJKimLHHanM Deep learning for pediatric posterior fossa tumor detection and classification: a multi-institutional study. AJNR Am J Neuroradiol. 2020;41:1718–25. 10.3174/ajnr.A6704 32816765 PMC7583118 PrinceEWWhelanRMirskyDMStenceNStaulcupSKlimoP Robust deep learning classification of adamantinomatous craniopharyngioma from limited preoperative radiographic images. Sci Rep. 2020;10:16885. 10.1038/s41598-020-73278-8 33037266 PMC7547020 AttallahO. MB-AI-His: histopathological diagnosis of pediatric medulloblastoma and its subtypes via AI. Diagnostics (Basel). 2021;11:359. 10.3390/diagnostics11020359 33672752 PMC7924641 AttallahO. CoMB-deep: composite deep learning-based pipeline for classifying childhood medulloblastoma and its classes. Front Neuroinform. 2021;15:663592. 10.3389/fninf.2021.663592 34122031 PMC8193683 JoshiPJalloGPereraRJ. In silico analysis of long non-coding RNAs in medulloblastoma and its subgroups. Neurobiol Dis. 2020;141:104873. 10.1016/j.nbd.2020.104873 32320737 Khayat KashaniHRAzhariSMoradiESamiiFMirahmadiMSTowfiqiA. Predictive value of blood markers in pediatric brain tumors using machine learning. Pediatr Neurosurg. 2022;57:323–32. 10.1159/000525862 35785770 JabarkheelRHoCSRodriguesAJJinMCParkerJJMensah-BrownK Rapid intraoperative diagnosis of pediatric brain tumors using Raman spectroscopy: a machine learning approach. Neurooncol Adv. 2022;4:vdac118. 10.1093/noajnl/vdac118 35919071 PMC9341441 DjirackorLHalldorssonSNiehusmannPLeskeHCapperDKuschelLP Intraoperative DNA methylation classification of brain tumors impacts neurosurgical strategy. Neurooncol Adv. 2021;3:vdab149. 10.1093/noajnl/vdab149 34729487 PMC8557693 BruschiMPetrettoACamaAPavanelloMBartolucciMMoranaG Potential biomarkers of childhood brain tumor identified by proteomics of cerebrospinal fluid from extraventricular drainage (EVD). Sci Rep. 2021;11:1818. 10.1038/s41598-020-80647-w 33469081 PMC7815722 EwejeFRBaoBWuJDalalDLiaoWHHeY Deep learning for classification of bone lesions on routine MRI. EBioMedicine. 2021;68:103402. 10.1016/j.ebiom.2021.103402 34098339 PMC8190437 HeYPanIBaoBHalseyKChangMLiuH Deep learning-based classification of primary bone tumors on radiographs: a preliminary study. EBioMedicine. 2020;62:103121. 10.1016/j.ebiom.2020.103121 33232868 PMC7689511 PanDLiuRZhengBYuanJZengHHeZ Using machine learning to unravel the value of radiographic features for the classification of bone tumors. Biomed Res Int. 2021;2021:8811056. 10.1155/2021/8811056 33791381 PMC7984886 WangHZhangJBaoSLiuJHouFHuangY Preoperative MRI-based radiomic machine-learning nomogram may accurately distinguish between benign and malignant soft-tissue lesions: a two-center study. J Magn Reson Imaging. 2020;52:873–82. 10.1002/jmri.27111 32112598 ConsalvoSHinterwimmerFNeumannJSteinbornMSalzmannMSeidlF Two-phase deep learning algorithm for detection and differentiation of Ewing sarcoma and acute osteomyelitis in paediatric radiographs. Anticancer Res. 2022;42:4371–80. 10.21873/anticanres.15937 36039445 YangYZhouYZhouCMaX. Novel computer aided diagnostic models on multimodality medical images to differentiate well differentiated liposarcomas from lipomas approached by deep learning methods. Orphanet J Rare Dis. 2022;17:158. 10.1186/s13023-022-02304-x 35392952 PMC8991509 ZhangXWangSRudzinskiERAgarwalSRongRBarkauskasDA Deep learning of rhabdomyosarcoma pathology images for classification and survival outcome prediction. Am J Pathol. 2022;192:917–25. 10.1016/j.ajpath.2022.03.011 35390316 FrankelAOLatharaMShawCYWogmonOJacksonJMClarkMM Machine learning for rhabdomyosarcoma histopathology. Mod Pathol. 2022;35:1193–1203. 10.1038/s41379-022-01075-x 35449398 PenederPStützAMSurdezDKrumbnholzMSemperSChicardM Multimodal analysis of cell-free DNA whole-genome sequencing for pediatric cancers with low mutational burden. Nat Commun. 2021;12:3230. 10.1038/s41467-021-23445-w 34050156 PMC8163828 KoelscheCSchrimpfDStichelDSillMSahmFReussDE Sarcoma classification by DNA methylation profiling. Nat Commun. 2021;12:498. 10.1038/s41467-020-20603-4 33479225 PMC7819999 LazovaRSmootKAndersonHPowellMJRosenbergASRongiolettiF Histopathology- guided mass spectrometry differentiates benign nevi from malignant melanoma. J Cutan Pathol. 2020;47:226–40. 10.1111/cup.13610 31697431 TognettiLBonechiSAndreiniPBianchiniMScarselliFCeveniniG A new deep learning approach integrated with clinical data for the dermoscopic differentiation of early melanomas from atypical nevi. J Dermatol Sci. 2021;101:115–22. 10.1016/j.jdermsci.2020.11.009 33358096 ZhangAJLindbergNChamlinSLHaggstromANManciniAJSiegelDH Development of an artificial intelligence algorithm for the diagnosis of infantile hemangiomas. Pediatr Dermatol. 2022;39:934–6. 10.1111/pde.15149 36164801 JiangCHLinPFChenFCChenJYXieWJLiM Metabolic profiling revealed prediction biomarkers for infantile hemangioma in umbilical cord blood sera: a prospective study. J Proteome Res. 2022;21:822–32. 10.1021/acs.jproteome.1c00430 34319108 O’BrienKFFriedenIJZeymoAVasicJSilvermanRGoldbergG Analysis of lesional color to differentiate infantile hemangiomas from port-wine birthmarks in infants less than 3 months old: a pilot study. Pediatr Dermatol. 2021;38:585–90. 10.1111/pde.14554 33742460 WuJZhouLGouFTanY. A residual fusion network for osteosarcoma MRI image segmentation in developing countries. Comput Intell Neurosci. 2022;2022:7285600. 10.1155/2022/7285600 35965771 PMC9365532 PengJKimDDPatelJBZengXHuangJChangK Deep learning-based automatic tumor burden assessment of pediatric high-grade gliomas, medulloblastomas, and other leptomeningeal seeding tumors. Neuro Oncol. 2022;24:289–99. 10.1093/neuonc/noab151 34174070 PMC8804897 StrijbisVIJde BloemeCMJansenRWKebiriHNguyenHGde JongMC Multi-view convolutional neural networks for automated ocular structure and tumor segmentation in retinoblastoma. Sci Rep. 2021;11:14590. 10.1038/s41598-021-93905-2 34272413 PMC8285489 BougetDPedersenAHosaineySAMVanelJSolheimOReinertsenI. Fast meningioma segmentation in T1-weighted magnetic resonance imaging volumes using a lightweight 3D deep learning architecture. J Med Imaging (Bellingham). 2021;8:024002. 10.1117/1.JMI.8.2.024002 33778095 PMC7995198 GiwaARossouwSCFataiAGamieldienJChristoffelsABendouH. Predicting amplification of MYCN using CpG methylation biomarkers in neuroblastoma. Future Oncol. 2021;17:4769–83. 10.2217/fon-2021-0522 34751044 ThurtleDRGreenbergDCLeeLSHuangHHPharoahPDGnanapragasamVJ. Individual prognosis at diagnosis in nonmetastatic prostate cancer: development and external validation of the PREDICT prostate multivariable model. PLoS Med. 2019;16:e1002758. 10.1371/journal.pmed.1002758 30860997 PMC6413892 FengHGuZYLiQLiuQHYangXYZhangJJ. Identification of significant genes with poor prognosis in ovarian cancer via bioinformatical analysis. J Ovarian Res. 2019;12:35. 10.1186/s13048-019-0508-2 31010415 PMC6477749 QianDLiuHWangXGeJLuoSPatz EFJr Potentially functional genetic variants in the complement-related immunity gene-set are associated with non-small cell lung cancer survival. Int J Cancer. 2019;144:1867–76. 10.1002/ijc.31896 30259978 PMC6377316 TopolEJ. High-performance medicine: the convergence of human and artificial intelligence. Nat Med. 2019;25:44–56. 10.1038/s41591-018-0300-7 30617339 HandelmanGSKokHKChandraRVRazaviAHLeeMJAsadiH. eDoctor: machine learning and the future of medicine. J Intern Med. 2018;284:603–19. 10.1111/joim.12822 30102808 YeungSDowningNLFei-FeiLMilsteinA. Bedside computer vision—moving artificial intelligence from driver assistance to patient safety. N Engl J Med. 2018;378:1271–3. 10.1056/NEJMp1716891 29617592 LarsonDBChenMCLungrenMPHalabiSSStenceNVLanglotzCP. Performance of a deep-learning neural network model in assessing skeletal maturity on pediatric hand radiographs. Radiology. 2018;287:313–22. 10.1148/radiol.2017170236 29095675 RatnerAJEhrenbergHRHussainZDunnmonJRéC. Learning to compose domain-specific transformations for data augmentation. Adv Neural Inf Process Syst. 2017;30:3239–49. 10.48550/arXiv.1709.01643 HavaeiMDavyAWarde-FarleyDBiardACourvilleABengioY Brain tumor segmentation with deep neural networks. Med Image Anal. 2017;35:18–31. 10.1016/j.media.2016.05.004 27310171