Introduction

Gliomas are the most common malignant tumors of the central nervous system (CNS), originating from glial cells such as astrocytes, oligodendrocytes, and ependymal cells1,2. Among these, glioblastomas represent the most aggressive subtype, with high recurrence rates and poor overall survival (OS). Despite advancements in treatment strategies, including surgery, radiation, and chemotherapy, the prognosis for glioblastoma patients remains dire, with a median OS of approximately 14.6 month3,4. The ability to accurately assess tumor progression and predict patient outcomes is essential for optimizing therapeutic approaches and advancing precision medicine.

Medical imaging plays a pivotal role in the diagnosis and management of gliomas, with magnetic resonance imaging (MRI) considered the gold standard for tumor visualization and treatment response assessment4,5. Advanced imaging techniques, including perfusion-weighted imaging, diffusion-weighted imaging, and positron emission tomography (PET), have further enhanced our understanding of tumor biology2. However, conventional radiological assessments often fail to capture the full complexity of gliomas, necessitating more advanced computational approaches.

Radiomics, an emerging field within precision oncology, leverages machine learning (ML) and artificial intelligence (AI) to extract high-dimensional qualitative features from medical images6. These features capture tumor heterogeneity, spatial distribution, and texture patterns that may not be discernible through traditional imaging analysis7. When integrated with clinical and molecular data, radiomics models can improve prognostic accuracy, facilitate early tumor progression detection, and support personalized treatment planning.

Despite the potential of AI-driven radiomics, challenges remain in standardized feature extraction, validating predictive models, and ensuring their reproducibility across diverse patient populations. The lack of consensus regarding optimal imaging protocols, segmentation methodologies, and model interpretability has hindered the widespread clinical implementation of these techniques. Addressing these limitations requires a comprehensive evaluation of existing AI-radiomics methodologies and a structured approach to model development and validation.

This study conducts a systematic review and epistemic meta-analysis to assess the current state of AI-radiomics integration in glioma prognosis. By analyzing methodologies, predictive features, and machine learning algorithms employed in previous research, we aim to identify best practices and propose a standardized roadmap for developing robust and clinically applicable AI-Radiomics models. Through this approach, we aim to bridge the gap between research and clinical application, ensuring that AI-Radiomics models are not only precise and reliable but also seamlessly integrable into existing glioma prognosis workflows.

Materials and methods

Article scaffolding

The global structure of the article is the following: The first step was formulating the research question, and the research objectives followed by separating the question into its components according to the PIO (Participants/Intervention/Outcome) strategy. This was continued by searching the keywords and MeSH (Medical Subject Headings) terms in three databases following PRISMA AND PICO approach8,9,10. Secondly, the resulting articles were assessed and reviewed. Thirdly, a epistemic meta-analysis was performed.

Search strategies

A systematic review was conducted until November 2023 using a modified PIO (Participants/Intervention/Outcome) approach, adhering to the Principles for Systematic Reviews and Meta-Analyses (PRISMA) guidelines to validate searches across various databases8,9,10.

PIO strategy

The search strategy was developed using MeSH terms and Boolean operators to address the research question. The primary objective was to explore the latest findings on the application of AI, particularly machine learning (ML), in predicting high-grade glioma progression using a radiomics approach. The PIO strategy encompassed Participants, Interventions, and Outcomes, focusing on the following research question: Which radiomic features are associated with progression in patients with high-grade gliomas assessed by MRI using AI methods such as ML?

The databases used were PubMed, BIREME, and Web of Science with the following configuration:

Participants Articles related to high-grade gliomas diagnosed by histopathology and MRI studies were included. Relevant terms included: Glioma, high-grade glioma, brain tumor, tumor microenvironment, magnetic resonance image, MRI, texture analysis, semiautomated, semiautomated analysis, semiautomated platform, biomarkers, precision medicine, and virtual biopsy.

Intervention Articles involving software and algorithms used to extract features from images were included, with terms such as: artificial intelligence, machine learning, Adaboost, Xgboost, Lightgbm, gaussian process, decision trees, gradient boosted decision trees, neural Networks, k-nearest neighbor, deep convolutional neural network, logistic regression, partial least squares discriminant analysis, quadratic discriminant analysis, random forest, stochastic gradient descent, support vector classification, fully connected network, Lifex, PyRadiomics, Olea sphere, 3DSlicer, Python, Python PyTorch, supervised learning and unsupervised learning.

Outcome Articles relating first and second-order radiomic features to the prognosis of high-grade gliomas were included.

First-order features Properties of individual voxels without considering their spatial relationships, such as intensity, histogram, conventional, and discretized.

Second-order features Analyzing the texture and spatial relationships between voxels within the region of interest:

  • Gray Level Co-occurrence Matrix (GLCM)

  • Gray Level Run Length Matrix (GLRLM)

  • Gray Level Size Zone Matrix (GLSZM)

  • Neighboring Gray-Tone Difference Matrix (NGTDM)

  • Gray Level Dependence Matrix (GLDM)

Shape and volume Geometric properties of the ROI.

  • Sphericity

  • Compactness

  • Elongation

  • Surface area

  • Size of ROI (volume)

Terms related to progression

  • Progression

  • Progression disease

To define the proportion of semiautomated or automated platforms used, we assessed the tumoral radiomic features in high-grade gliomas and their association with the tumor microenvironment (edema). Additionally, we evaluated the models proposed to predict progression using radiomic analysis through AI.

The following questions were proposed to guide the study; what proportion of semi-automated platforms assess radiomic features in high-grade gliomas? What percentage of automated platforms assess radiomics features in high-grade gliomas? Which radiomic features have been associated with the tumor microenvironment in high-grade gliomas? and, what are the main methods of modeling and validation of proposed AI-based algorithms?

Databases and searches

The electronic databases used were PubMed, BIREME, and Web of Science. The search strategy was replicated for each database using the combination of several Booleans appropriated for our research question (Fig. 2). This figure describes in full detail the search strategy in the PubMed database as well as all keywords used. Subsequently, the results obtained from these searches were recorded in the Mendeley database.

IA application to predict the progression of disease

This systematic review followed the PRISMA strategy to assess the articles pertinent to the research question.

Quality control assessment

A quality control has been carried out taking into consideration the methodological aspects of the articles obtained for the analysis.

We used the criteria assessment proposed by Juarez et al.7 that briefly encompasses:

  1. 1.

    Research questions and objectives. A clear question, a clear objective.

  2. 2.

    Methodology: Complete, robust, and adequate; Suitable type of study, suitable instrument, reproducibility, and bias control.

  3. 3.

    Results: Congruent with the objective, congruent to the question.

  4. 4.

    Terms: Properly defined.

  5. 5.

    Conclusions: That answers the question and achieves objectives.

The quality control score ranges from 0 to 100, with articles scoring 75 or higher being accepted for the meta-analysis. A total of 19 articles were assessed for quality, resulting in their inclusion in the final meta-analysis.

Integral epistemic meta-analysis: exploring the application of AI in the progression of high-grade gliomas

Epistemic meta-analysis is an approach used in research to analyze and synthesize existing knowledge, theories, or findings. On this research, the term ‘epistemic’ refers to the examination of knowledge or understanding by evaluating the underlying assumptions, methodologies, and theoretical frameworks of AIRI studies. Unlike a traditional meta-analysis which is only quantitative and requires data that is homogeneous, consistent and unified between studies, an epistemic meta-analysis allows to gather knowledge from articles that are not easily comparable, which is the case of the AIRI field to this date because each paper reports results with different metrics and the studies had to be divided into four topic groups to be comparable. The precedent situation requires to first create a puzzled construction by a epistemic meta-analysis, which is the goal of this study. Additionally, this approach aims to provide a deeper understanding of the current state of knowledge within the AIRI glioma domain, highlighting its strengths, limitations, and areas for further investigation. The objective is to create a roadmap and guidelines that help to homogenize and integrate the results of the field, so that in the future, the field can be more robust for traditional meta-analysis to be viable.

This study was conducted by a cross-functional group consisting of medical doctors, oncologists, radiomics experts, a physical engineer with experience in AI, and bioinformaticians.

Integral AI meta-analysis algorithm (IAIMA)

The 19 studies selected in the systematic review that passed the quality control assessment were explored with an IAIM. The Integral AI meta-analysis algorithm is a process where the common elements of the article scaffold are identified, analyzed, and classified. It is a process that encompasses 5 steps. First, the studies were classified by application topics. This classification is shown in Fig. 4. The topics are prognosis of OS or PFS, progression after treatment, segmentation, and other stratifications. Second, all the relevant data (clinical, radiomics, processes, components, etc.) for building the model was registered in mind maps, starting from the common components of the scaffold and identifying the variable components (Fig. 2). Third, the data gathered led to building two organized tables consisting of what was done in the studies (Table 2) and how it was done (Table 3). Fourth, the studies were analyzed by two methods (mixed method, and by application topic) based on Tables 2 and 3. Fifth, a proposal of a pipeline to follow for the development of an enhanced model for the main application topics that could be implemented in hospitals was done (Fig. 8).

This approach made it possible to compare the studies in relatively equal terms. This meta-analysis served the purpose of gathering the key elements for building an AI radiomics model for the prognosis of OS, prognosis of PFS, and progression after treatment. The combination of the best scaffolding of each article is the starting point to make a pipeline for building a model that would help oncologists and doctors perform a kind of triage for patients’ care centers.

Expanded algorithm of the meta-analysis

To visualize all the processes, we describe each step mentioned above.

Application topic classification

The studies were classified by application topic. The classification of topics helps visualize the different models that can be developed to help the innovation and treatment of cancer patients with high-grade gliomas. This was done by a thorough reading of the papers. Some studies had two topics, so a Venn diagram was used to represent the classification. There are 4 topics: prognosis of OS or PFS, progression after treatment, segmentation, and other classifications. The prognosis of OS or PFS means that the article’s model stratifies patients regarding their prognosis of OS or PFS. The topic of progression after treatment consists either of articles that diagnose the patients’ progression condition (pseudo-progression or true progression) some weeks after the treatment, or articles that predict the future progression areas outside the tumor from the pre-operative MRI sequences. Segmentation articles focus only on the division of the regions of the tumor and report that performance. The classification topic means that a model from the article classified something else than prognosis, progression after treatment, or segmentation. These articles classified either the REP condition, the GBM molecular subtype, or the molecular biomarkers, glioma grade, and identified progression from no progression.

Studies’ roadmaps

The steps of the roadmap of each article were identified as computable objects. This means that the input of the system is complex and integrated with clinical, molecular, and image data. Followed by pre-processing, segmentation, and feature extraction of the images that are then converted into numerical values named radiomic features. The roadmap is the series of steps that each study did to develop their model. According to the scaffolding of the articles, a general roadmap of what most studies did was identified.

Mind maps

The mind maps served as a tool for summarizing the key information of the development process of each model and its performance. The data of both the best-performing model and additional models built on each article were identified, registered, and classified. A baseline scaffolding-template mind map was used for all the studies to standardize both the structure and the data summarized in this step (Fig. 1). The mind maps allow a deep understanding of each of the nineteen articles in a glimpse, and they are also effective for looking up and overlapping specific and relevant data of the study. They are also helpful in identifying the elements of the complete model development process.

Fig. 1
figure 1

Base-line mind map for step 2 of the meta-analysis.

Full size image

The mind maps consist of two parts: what was done (right), and how it was done (left). The right side has the main information of the article, while the left side has the information that supports the results of the article, it summarizes how it was done (Fig. 1).

In Fig. 1, the right side has the following branches: (i) identifying the application topic or type of article, (ii) the objective, (iii) the cancer type, (iv) the model and its performance, and (v) additional specific results of the article. (i) The applications of binomial AIRI of the articles have four main categories: prognosis of OS or PFS, progression after treatment, segmentation, and other classifications. Some articles have two application topics. (ii) The objective of the article is the purpose of why it was written. (iii) The cancer type was divided into three: GBM, LGG, and DMG. (iv) The performance of the model was reported with different variables, depending on which variables were used in the article. (v) Some articles performed additional studies that diagnose: REP condition, GBM molecular subtype, molecular biomarkers, or glioma grade.

In Fig. 1 the left side, which consists of how the study was carried on, has two proper branches, which are (i) data and (ii) approach, and one branch that is common with the right side, (iii) the model branch. The following elements were included on the (i) data branch: imaging types, number of patients, source, availability of the parameters used for the MRI sequences, and the pre-processing information. The (ii) approach branch has two sub-branches: features and segmentation. The features sub-branch has registered the number and type of features used, the feature selection process, and the software used to extract the radiomic features. The segmentation sub-branch has the segmented regions, the level of automatization for this process, and the software used.

AI Table: information for the creation of the model

The information gathered in the mind maps was then organized on a table. It was placed in such a way that it gives the possibility to compare any article with the others on any aspect. The data of the best-performing model of each application topic was the one registered on Tables 2 and 3. In Chilaca’s study, rather than training an AIRI model, the researchers focused on evaluating the predictability of individual features without creating a comprehensive model. This alternative approach is particularly appealing for the oncologist, for it emphasizes the relevance of each feature. The features’ performances reported for that study on Table 2 were CONV_RIM_stdev for OS and GLCM contrVar for PFS12. The AI table has six subheadings, which are: Main Information, Model Information, Model Performance, Segmentation, Features, and Data. The AI table is divided into Tables 2 and 3. The division criterion is the same as on the mind maps (what, how).

Table 2 has three headings; main information, model information, and model performance. The main information heading has general information about the study, the topics covered in it, the clinical data, and the study’s objective. The model information subheading has the specifications of the machine learning model, it consists of 7 columns. The model performance heading contains how well the model did and it has both the training and testing performances registered. The performance variables used were AUC, accuracy, sensitivity, specificity, and others (e.x. p values). If any of these variables was not reported, it was specified as blank (–).

Table 3 has three headings as well: segmentation, features, and data. The segmentation heading consists of segmented regions, the number of regions, the level of automatization, and the software used for the segmentation process. The features subheading has 7 columns. It registers which features were used, with which software they were extracted, how the feature selection process was done, and which ones were the most predictive features. Last of all, the data branch has the MRI sequences used, the status of the patient’s treatment upon imaging, the number of patients, the split of training and testing, and the sources of the data.

Model comparison

Two types of analysis were carried on, a mixed method analysis (quantitative and qualitative), and an analysis by application topic. The mixed method consisted of analyzing together all the articles on a specific aspect. There were three aspects analyzed: clinical patient data, radiomic features and segmentation, and ML models. This led to finding the different proportions that answer some of the research questions. On the other hand, the analysis by application topic consisted of first grouping the articles in the topics previously identified in step 1 of the meta-analysis and then comparing the articles within each application topic separately. This analysis made it possible to compare the performance of the models against each other. The application topics analyzed were two; prognosis of OS or PFS, and diagnosis progression after treatment. The AI table served as the source material for making the comparisons. The models with the highest performance scores were identified and highlighted. The features from these models were considered to be the most predictive ones, taking into account that the features used to train a model are the key elements for the model’s performance. Furthermore, the best algorithm to use and a robust methodology to develop a model were explored.

Model pipeline proposal

A proposal for a pipeline to follow to develop an AIRI prognosis model and its implementation in a clinical setting was done. The development approach combines the methods, features, and ML algorithms used on the best-performing models. The development proposal is a roadmap with eight steps with specific suggestions to follow.

Results

State of the art of progression in high-grade gliomas using AIRI

The research question of this systematic review: Which are the radiomic features associated with progression in patients with high-grade glial tumors assessed by MRI using AI methods such as ML. The PIO dissected the research question in these parts: Participants with histopathology diagnosis of high-grade glioma; Intervention assessed with an MRI analysis by semiautomated radiomic platform; Outcome features associated with progression disease. We can follow the PIO strategy in Fig. 2.

Fig. 2
figure 2

A representative decision diagram showing the articles retrieved using the PIO strategy in the PubMed database. P (participants). I (interventions). O (Outcomes).

Full size image

PRISMA assessment results

This systematic review began with 210 references obtained from three electronic databases: 79 were from PubMed, 69 were from BIREME, and 63 were from Web of Science. In the first screening, double references were excluded, removing 105 articles.

A second screening (eligibility) was performed using the following set of quality criteria:

  1. 1.

    Original articles (not review articles, letters, abstracts, or viewpoints).

  2. 2.

    Published in 10 years.

  3. 3.

    Written in English.

These criteria were applied, leaving with 70 articles.

A third screening was conducted to exclude articles that were irrelevant to the review, resulting in 51 exclusions and leaving 19 articles for the review. See Fig. 3 and Table 1.

Fig. 3
figure 3

Identification of studies via databases and registers with the PRISMA approach.

Full size image
Table 1 Quality control assessment.
Full size table

Quality control results

The results of the quality control are described in Table 1, which integrates all the data on the criteria used to qualify the methodology of the studies included for the systematic review.

Meta-analysis results: AIRI binomial as a route to improve diagnosis of patients’ progression

Application topics of studies

As for the application topics of the nineteen selected articles, eleven were prognosis of OS or PFS, six were progressions after treatment, and two were segmentation. Three articles whose topic was the prognosis of OS or PFS also did another type of classification. A diagram with the classification of the study’s application topics is shown in Fig. 4.

Fig. 4
figure 4

Application topics of systematic review.

Full size image

General AIRI roadmap

Most of the articles share common steps in their AIRI roadmaps. First, the patient selection is done, consisting of MRI, clinical, and molecular data acquisition. Then the MRI is pre-processed followed by the segmentation of the tumor, where the regions of interest (ROI) are identified for each image. Afterwards, the radiomic features are extracted from the ROI of the images. Then feature selection is performed, where the most predictive features are chosen (radiomic, clinical, and molecular). This is followed up by training the machine learning or statical model, where a mathematical model is built with the most predictive features and a particular algorithm. Finally, the performance of the model is assessed for both the patients with whom the model was developed (training set) and for patients who were left out of the model’s development process (testing set). Sometimes feature selection, training the model, and assessing its’ performance work as a single cyclic step (example: Recursive Feature Elimination (RFE)15,21, Backward Feature Elimination (BFE)13). This general AIRI roadmap is shown in Fig. 5.

Fig. 5
figure 5

AIRI roadmap (pipeline). The steps involved in the radiomics analysis workflow start with image acquisition and pre-processing and end with the application and evaluation of the outcome. After pre-processing, segmentation of the tumor regions (ROI) in MRI sequences is performed to then extract data from the segmented regions (feature extraction). Afterward, the most predictive features are selected (feature selection), and the ML model is trained with a specific algorithm (model training). AI has been incorporated to perform predictions of outcomes as the final step.

Full size image

Mind maps of the AIRI epistemic meta-analysis

The baseline scaffolding-template mind map proved useful to analyze all the articles disregarding the differences in application topics. Despite that, new sub-branches were added or branches were trimmed to adequate the baseline mind map for each article. When no information was found on a specific topic, the letters N.S. (not specified) were used. Figures 6 and 7 show two example mind maps of articles with different application topics.

Fig. 6
figure 6

Example mind map of an article on the prognosis of the overall survival (OS) or progression-free survival (PFS).

Full size image
Fig. 7
figure 7

Example mind map of an article on progression after treatment.

Full size image

Chang et al.25 study was taken as an example of the mind map analysis (Fig. 6). On the first visualization, one can see that the right side corresponds to what Chang study did, and the left side about how he did it. Subsequently, the mind map is divided into four quadrants. In the upper right quadrant are all the study characteristics (orange, purple, and pink). In the upper left quadrant are all the clinical features and the process to obtain them (green). In the lower left quadrant is the base data of the model (blue). In the lower right quadrant is the data related to the machine learning algorithm and its’ performance (yellow). This mind map structure is common to all studies. Overlapping the minds maps allows to identify on the fly the constant parts of the studies and find their differences based on their application topics.

Patel et al.15 study was taken as another example of the mind map analysis (Fig. 7). The constant parts of this mental map compared with Chang’s study, are: That the right side corresponds to what they did, and the left side to how they did it. The upper right quadrant shares most of the structure, but the content adapts to the application topic. The upper left quadrant stays the same. In the lower right quadrant, both pink branches are of statistical analysis, and the yellow branch basic structure remains relatively constant. The lower left quadrant stays the same. The parts that differ from Chang’s study are the type of article on the upper right, the specific approach of the statistical analysis of the pink branch, and the way the model’s performance is reported on the yellow branch due to Chang developing four different models. This comparison shows that this mind map analysis allows to identify on the fly each studies’ strengths, methods, and specific approach, recuperating the essential parts of each research.

AIRI meta-analysis table

In the tables (Tables 2, 3), when a particular article created models for two different application topics, both were reported. Sometimes camel case convention was used to optimize space. To organize information, the following convention was used: category:{element1,element2,…, element n}. On some cases, the total number was specified inside brackets.

Table 2 Summarized information on the development of the AIRI models of the 19 studies (What?).
Full size table
Table 3 Summarized information on the building of the model of the 19 articles (How?).
Full size table

Finding the cornerstone of AIRI: comparison between studies, a mixed method approach.

Clinical characteristics in high-grade glioma

From the evaluated studies, a striking 94.4% are centered on adult patients (Table 2), while there has been less investigation on the pediatric population. This is reasonable because incidence of malignant glial tumors in the pediatric population is lower, ranging from 3 to 15%28. However, the prognosis for primary central nervous system tumors in children remains equally grim as in adults, underscoring the importance of exploring this field. Some variances have been noted in imaging assessments, particularly in differentiating characteristics observed in T2 multiparametric resonance sequences29. Thus, evaluating radiomic features across T1 sequences with contrast, T2, FLAIR pre- and post-treatment becomes imperative, shedding light on the distinct tumor habitat behaviors in pediatric glial tumors (Table 3).

While most studies concentrate on hemispheric lesions, our previous publication12 uniquely addresses midline tumors, recognized as the most aggressive subgroup. This pathology merits greater global attention due to its particularly grim prognosis. Thus, emphasizing efforts in understanding and addressing midline tumors could significantly impact overall outcomes in glioma management.

Segmentation and radiomics characteristics

The first finding was the heterogeneity of the methodology applied in radiomics processes (identification of volumes of interest, segmentation, and analysis) with AI support in neuro-oncology.

In radiomics analysis, most studies extracted first and second-order features, with a mean of 2,087 initial features per patient (Table 3). These features were obtained based on various analysis parameters, including regions of interest, number of segmentations, and image sequences. Given the large volume of data, we emphasize the importance of utilizing AI and big data analysis platforms to manage and interpret the results effectively. Many studies did not include clinical features, while some incorporated molecular features (58%) to help avoid bias (Tables 2, 3).

Second-order radiomic characteristics were used in 68% of the studies. The alterations on these features provide detailed metrics on the tumor microhabitat, enhancing our understanding of the diagnosis and impacting surgical treatment plans as well as other therapies such as radiotherapy. This allows for precision and personalized treatment tailored to the specific phase of the oncological disease in the central nervous system.

In the segmentation step, 52.6% of studies focused on both the tumor and other regions, while 36.8% focused solely on the tumor. Managing multiple regions can complicate the analysis of the extensive data obtained from each segment. Therefore, it is imperative to reach a consensus on the regions of interest (ROI) to be segmented and to establish the minimum necessary volumes for ROI, such as tumor, edema, and necrosis for prognosis models, and only tumor for progression after treatment models, which are the number used on the best performing models (Tables 2, 3). Particularly, tumor and edema are crucial for clinical and therapeutic planning, including surgery and radiotherapy. The lack of consensus on the number of regions to segment, the methodology to follow, and the accuracy of results presents significant drawbacks, leading to discrepancies in radiomic acquisitions and analysis. Additionally, the software used to extract features varied between studies, with Pyradiomics and MATLAB being the most commonly used (on four studies each) (Tables 2, 3).

Characteristics of the artificial intelligence models proposed in the studies

Most studies (76%) employed machine learning algorithms, while only a smaller portion (26%) utilized statistical algorithms (Table 2). The most commonly used machine learning algorithms were Random Forest (RF), featured in six studies, and Support Vector Machine (SVM), implemented in three studies. LASSO, combined with either linear regression or logistic regression, was also used in three studies. Other algorithms included Linear Discriminant Analysis (LDA), Bayesian Networks (BN), CatBoost (CB), and two types of neural networks (NN) (Table 2).

Finding the cornerstone of AIRI: comparison between studies, following the footsteps of the best binomial AIRI models—application topics analysis

Prognosis of OS or PFS studies

Regarding the studies that did prognosis of OS or PFS, the ones that had the best performance were: Su-2021, Chaddad-2018, Pak-2018, and Sanghani-2018. For predicting two class OS, the k-fold cross-validation test scores of the top articles were an AUC of 85.5%20, and an accuracy of 98.7%21. The mean AUC and accuracy of the studies that predicted two class OS were 83.2% with a standard deviation of 11.6% and 87.3% with a standard deviation of 11.2% respectively. For the three class OS classification, the top test scores for the k-fold cross-validation were an AUC of 98%13 with an accuracy of 94.8%13 and an accuracy of 89.0%21. For two class PFS, the best k-fold validation test scores were an AUC of 85.4%20 and a p value on the long rank test of < 0.00116. All this information can be found in Tables 2, 3.

The features that proved to be predictive for the prognosis of OS or PFS were many. Because the initial features varied from study to study, the features found to be the most predictive after feature selection differed from author to author. The standard features (features that were included in most studies) that proved predictive are 1st order (kurtosis, energy necrotic Ve map), standard deviation, as well as others), 2nd order (GLCM, GLDM, NGTDM, GLRLM, GLSZM), shape and volume, and clinical and molecular features (age, IDH, and MGTM). There were also innovative features that proved predictive in specific articles. For example, logFilter texture features were the only features used in Chaddad’s model, which had a good performance. Gabor texture features addition to the standard features were used in Sanghani’s very good-performing model. Su’s innovative model used only deep imaging features (4D map, 2D fisher vectors).

The algorithms behind the models that had the best performance for the prognosis of OS or PFS were SVM, RF, and LASSO. The two models that used the SVM algorithm were among the top-performing models13,21challenging to generalize which is the best algorithm to use. It was not a univariate analysis where only the algorithm changes. The features and the methods used in each article varied additionally to the algorithm used. Despite that being the case across articles, some indeed performed a univariate analysis of the algorithms. These articles had the profitable approach of training several models, each with a different algorithm while maintaining the other variables constant, and then picking the best-performing one. For example, Su trained three models with different algorithms; SVM, RF, and LR, and reported accuracies of 94.8%, 69.6%, and 84.6% respectively13. In this case, the SVM algorithm outperformed the others.

Progression after treatment

There were six studies whose application topic was progression after treatment, but only four of those diagnosed patients between presenting pseudo-progression from tumor early tumor progression (which excludes Yan’s and Zhang’s studies). Zhang’s study classified patients between presenting injury from treatment from recurrence, and Yan’s study predicted future progression areas. The four previously mentioned articles and Zhang’s study were included in this analysis because they all diagnose the state of progression after treatment.

The models that have the best performance are Zhang’s-2022, Kebir’s-2020, and Kim’s-2018. For models that diagnosed pseudo-progression from true early progression, the top test scores for the held-out test set are AUC of 93%22 and 85%19. For models that classified injury against recurrence, Zhang got scores of 94% AUC, 94.4% sensitivity, and 91.7% specificity on the held-out test set31. All this information can be found in Tables 2, 3.

The features that proved to be predictive for progression after treatment were more noticeable to identify than those for prognosis. This was because four out of the five articles compared use similar features. These features were: 1st order, texture, shape plus volume, and wavelet transform features of the previously mentioned ones. Despite that, the only features used in Kebir’s study were PET features (TBRmax, TBRmean, TTP), and the model performed well22.

There are different ML and statistics algorithms in the studies that performed progression after treatment. The above good-performing models used SVM31, Linear Discriminant Analysis (LDA)22, LASSO Logistic Regression19. Zhang used the strategy of training four models with different algorithms and then highlighted SVM as the best one31.

Proposal of a pipeline to develop and implement an AIRI model

The outcome of the systematic review and AIRI meta-analysis led to the proposal of the following two-fold AIRI universal roadmap.

Fold one: AIRI universal roadmap development

(1) MRI: For optimum performance, all the MRIs should be taken with the same scanners and with the same parameters such as magnetic field intensity (T), caption time frame, matrix size, number of excitations, slice thickness, gap, and equipment brand14,19. This hold true except if the model will be used to diagnose patients with images that will be taken with scanners or parameters different from the ones used to train the model. In that case, it would be better to use MRI from a variety of different equipment and parameters to train the model, or best, it would be to use MRI from all the equipment and parameters that are going to be used to diagnose patients. This would be the case if the model is intended to be used in several medical centers that have MRI equipment that differs. It is important to keep this point in mind because one could incur in selection bias if one does it improperly.

The MRI imaging types that are suggested to be incorporated into the model are T1W, CE-T1WI, T2W, and FLAIR. Some studies only used CE-T1W, and T2 or FLAIR and also achieved good performances. According to the Brain Tumor Segmentation competition (BraTS) CE-T1W1 compared against T1, and FLAIR are key for the automatic segmentation of the tumor30.

(2) Image pre-processing: It’s important to do image pre-processing before segmentation. Some general steps are normalization, interpolation (to achieve the same high resolution on all the MRI), removing outlier voxels, registration, and to skull strip the MRI. Some available software is 3D Slicer, PyRadiomics15, FSL27, Matlab—SPM plug-in25,31, ITK; N4ITK; ITK-SNAP 15,24, R—ANTsR and WhiteStripe pakages19, FMRIB18, NordiIce18. It would be ideal to automatize this step with a specific development so that no user needs to interact and do this process.

(3) Segmentation: It is recommended to segment the tumor into three regions, as the studies which used three regions for prognosis had the highest performances (Tables 2, 3). These regions are tumor, edema, and necrosis. On the studies that did treatment progression, most good performing articles only segmented the tumor in one region, tumor from no tumor, so that should be sufficient for that application topic. Once segmented, the segmentation masks should be applied to all the imaging types. For it to be feasible to implement the model and diagnose patients, the segmentation process should ideally be automatized16. Available software for doing this process automatically are 3D U-Net32, and RA-Unet13. 3D U-Net was used to segment into three regions (edema, tumor, necrosis), only using as input CE-T1 MRI sequences32. This makes it very practical, although the accuracy of this method might be lower than the semi-automated or manual approach (refer to other available software’s on Table 3). RA-Unet on the other hand, needs all four T1WI, CE-T1WI, T2 and FLAIR, and is not a conventional segmentation method, it must be used with Su’s publicly available prognosis model workframe13. All the other software’s used on the studies were semi-automatic or manual. For it to be viable to implement the model to diagnose patients, it is recommended that the automatic approach is chosen because the others are time consuming16. The current state of automatic software’s is rudimentary, and usually radiologist do it better. However, if a neural network is created with good accuracy, it would be a steppingstone for automatizing and diagnosing patients on medical centers. The BraTS challenge is an annual competition where researchers are trying to do just that. Researchers from all over the world compete to develop an automatic model that does segmentation-related tasks.

(4) Features extraction: It is suggested to extract the following standard radiomic features: 1st order (all), 2nd order—Gray-Level Texture Matrixes (GLCM, GLDM, NGTDM, GLRLM, GLSZM) shape and volume features, and clinical and molecular features (age, functional status and molecular status in respect to IDH, MGTM). This can be done in Python with the package PyRadiomics, or in MatLab. Additionally, it is advisable to include innovative features, such as texture LoGFilter first order features, and Garbor textures, which can help to improve the model’s performance. Studies whose application topic is progression after treatment, also included wavelet transform features, achieving good results. Therefore, if this type of model is going to be trained, it is recommended to include those types of features. It is recommended to automatize steps two to four on a single software development so that later, it can be used on the implementation phase as well (Fig. 8).

Fig. 8
figure 8

Model’s Development and Implementation Roadmap. Top part, model development. Middle part, implementation process. Bottom part, patient’s clinical natural history. This figure is exemplified with the model algorithm behind being LASSO.

Full size image

(5) Split of training and testing data: It is important to separate the testing data from the data that will be used for feature selection and model training. If one does not leave apart the data for testing before doing feature selection, there is a risk that the features that will be selected are the ones that best describe the testing set. Separating both sets from the beginning, ensure that the reported testing performance is the one to expect when the model is implemented and avoid bias. In a way, this could be termed confirmation bias because one is only looking at the features that one already knows will correctly describe the testing set. It is as knowing ahead of time which features will work best for the particular testing set.

(6) Feature Selection and Model training: Best performing models combined feature selection with model training13,21 using either RFE or BFE. This process can be done by evaluating the training performance alone, or computationally more time-consuming but better, is to do it with k-fold cross-validation21. In RFE and BFE, the number and combination of features that achieve the highest performance are selected. To make a more robust model, proof against overfitting, it is best to choose less than that number, which usually has many features. To reduce computational time, it is recommended to do Pearson Correlation coefficients before doing RFE or BFE to eliminate highly correlated features. Many articles did this as a first step for feature selection.

(7) Choosing the best model: For best results, it is a good approach to train several models with different algorithms, for example, SVM, LR, and RF13, and then identify the best-performing model13. Performing k-fold cross-validation for each model is highly recommended, if not already done (on the feature selection step), to evaluate the model’s performance.

(8) Testing the model:

  • Input: Use unknown new data (held-out test set from step 5).

  • Prediction: Make predictions with the trained model.

  • Performance Indicators: Obtain all performance indicators (AUC, accuracy, specificity, and sensitivity).

  • Interpretation: Check for overfitting and general outcome.

  • Reporting: Report all performance indicators in order to ensure transparency and foster comparability between studies15.

Testing the model with this dataset ensures the reported performance is trustworthy, making it a reliable expectation for clinical tasks such as prognosis, diagnosing progression after treatment, and other classifications. To check for overfitting, compare the test score with the training score. If the training score is significantly higher, the model is overfitted. One way to address overfitting, which is crucial to report, is to reduce the number of features. Selecting fewer features than the optimum number obtained in step 6 using RFE or BFE can help make the model more robust, universal, and less overfitted. If this adjustment is made and a new model with fewer features is trained, it is essential to note this in the report. If the model is retrained with less features and this is not reported, one is leading to a biased test performance indicator because the new model will be optimized against the testing set, by selecting the features that at the end describe the testing set best. This type of bias has been previously explored on step 5 of this section.

However, further adjusting the model to improve the held-out test set performance may compromise the validity of the held-out test set. This is because optimizing the model against the held-out test set effectively creates a model tailored to that specific test set, thereby nullifying its use as an unbiased test.

Fold two: AIRI universal roadmap implementation

It is suggested that when the model is developed, the development process be divided into two software’s, as exemplified in Fig. 8, and leave the performance analysis as another separate step from those software’s. Additionally, to this two software’s, a third software needs to be developed to get the patient data, apply the first and second previously developed software’s, and present both the classification result and the reasons behind it to the oncologist.

(1) A window where the patient’s MRI and clinical data is loaded must be developed. Another option is that the MRI and clinical data be automatically fed to this software.

(2) After that, the previously developed software on steps two to four which automates the image preprocessing, ROI segmentation and does feature extraction should be implemented using those MRI and clinical data as input. This process output is the radiomic and clinical features of the patient. To have an efficient data storage, it is advisable that the radiomic features that are not part of the feature selection cohort be eliminated or not stored on the database.

(3) After having obtained the radiomic and clinical features of the patient, the machine learning model trained in the development process should be applied to predict or classify the patient’s condition, taking as an input those features. The output of this process is the classification of the patient.

(4) The output of the classification should be presented on a visual manner together with the explanation as to why the model predicted that output. It is important to present this explanation, as this will aid the oncologist to understand the process behind it and therefore allow him to take better decisions. The explanation is also important so that the oncologist can tell the patient why he was classified in such a way.

Discussion

Epistemic meta-analysis is an approach used in research to analyze and synthesize existing knowledge, theories, or findings. On this research, the term ‘epistemic’ refers to the examination of knowledge or understanding by evaluating the underlying assumptions, methodologies, and theoretical frameworks of AIRI studies. It allows to understand the relevance of several types of features and to see how they get integrated into AIRI models. This approach generates the creation of ad hoc mental maps in order to identify the common backbone of the research trajectories that are common in all the investigations in the AIRI and distinguish them from the variables that are specific of the backbone of each study. This is invaluable as it allows for the creation of a navigable roadmap that can evolve. As it evolves, it can acquire new properties that enable a more generalized interpretation of results and open the epistemic horizon. Beyond academia this aids oncologist for the decision making in the data interpretation, diagnostic, prognostic, treatment, and follow up for the best care of the patients by giving insights on how to develop and implement AIRI models.

This study aims to unify the AIRI field and to expand the epistemic horizon so that in the future, a robust quantitative meta-analysis with meta-analytic techniques can be carried out. This study demonstrates and raises awareness to the scientific community the lack of comparability and heterogeneity of the methods, processes, and metrics used in the AIRI field, showed in the “AIRI meta-analysis Table” section (Table 2—What?, Table 3—How?), which was done meticulously. A general roadmap was presented as a suggested guideline for scientists to follow on future research, so that the field can evolve into a more robust one.

In this research, the expansive potential of the AIRI in high-grade glial tumors, primarily glioblastoma was explored. Swift diagnoses are imperative for conditions such as high-grade gliomas, given their bleak prognosis. The intricate journey from diagnosis to treatment and progression evaluation underscores the importance of specialized oncology reference centers with health care professionals training and becoming familiar with the AIRI. The promotion of integrating automation technologies for capturing, identifying, and discriminating between tumor characteristics, habitat, and surrounding healthy tissue in digital images of magnetic resonance studies is crucial. This integration facilitates precise diagnosis and treatment planning customized to meet each patient’s specific needs, thereby minimizing toxicity to healthy tissues in individuals with brain tumors. It encompasses the consideration of personalized-integral and effective management strategies, particularly pertinent in challenging neoplasms such as diffuse midline gliomas.

Towards a standard for AIRI interfiled: good practice

There must be a consensus as to how training and testing data is handled. Many studies have certain bias in their reported performances. Some models16,20 are particularly robust compared to others13,21. The robust ones, leave the test set completely separated from the process of developing the model, as shown it Table 2 on the column “held-out test set form fe. se.”. On the other hand, some studies performed feature selection (RFE and BFE) with the data that was used for testing on the k-fold cross-validation, and have no further external validation data. This approach makes the model susceptible to overfitting, giving the possibility for the model’s performance to drop when tested with real-world data. This scenario could happen because the features selected to build the model are those that describe the testing set particularly well.

Several clinical characteristics are excluded in different proposed models due to their perceived potential bias and as covariates that may alter the direction of progression prediction results. However, few studies have assessed stratification by molecular identity as a means to mitigate bias.

Table 2 illustrates that only 57.6% of the studies evaluate molecular alterations, indicating an area for improvement in stratifying these molecular alterations alongside clinical data and functional status.

The Radiomics Quality Score (RQS) was not utilized in these studies. RQS serves as a valuable tool aimed at enhancing the quality of radiomics research and necessitates widespread acceptance. It employs a system of rewards and penalties to evaluate the methodology, analysis, and reporting, thereby fostering optimal performance in radiomics analysis and scientific practice. While not yet established as the definitive standard for evaluating radiomics studies, efforts have been made to develop two versions of RQS, with the latest iteration (RQS 2.0) currently under development. Both versions comprise a comprehensive questionnaire consisting of 36 checkpoints, focusing on principles such as Fairness, Universality, Traceability, Usability, and Robustness, elucidating the efficacy of radiomics analysis.

It is also necessary that all the studies calculate and report all the model’s performance variables. Most studies reported the AUC score, but not all of them (Tables 2, 3). Some others reported the accuracy, but many did not. The same happened for specificity and sensitivity. It is necessary that all studies report all the machine learning performance indicators: AUC, accuracy, sensitivity, and specificity to ensure transparency of the results. It is straightforward to calculate them. It is advisable to show all the results and not only the best ones.

Excluding the testing set from the feature selection process, reporting all of the model’s performance indicators, including clinical and molecular features, and using RQS would enhance the comparability of studies, enabling more conclusive analyses. This approach would expedite progress in the AIRI by facilitating the identification of optimal strategies regarding feature selection, algorithms, MRI sequences, and ROI.

This work paved the way with a roadmap and methodology (Figs. 1, 6, 7, Tables 2, 3) and opened a new epistemic horizon constructed with the effort of all the studies included on this research11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,31,32 AIRI is now an interdisciplinary field, promoting collaboration and the creation of new projects where information is obtained and integrated universally, allowing research to be strengthened and fostered among different groups worldwide, sharing their knowledge as tools, ends, and practices. Thus, advancing AIRI in a joint direction that encourages rapid evolution in this field.

Conclusion

The integration of Artificial Intelligence and Radiomics Inter-field (AIRI) in neuro-oncology, particularly for predicting high-grade glioma progression, represents a paradigm shift in precision medicine. This systematic review and epistemic meta-analysis highlight the strengths and limitations of existing AIRI models while proposing a structured road map to improve standardization, reproducibility, and clinical implementation (Figs. 1, 2, 3, 4, 5, 6, 7, 8, Tables 2, 3).

Key findings indicate that robust radiomic features, particularly first- and second-order texture metrics, combined with machine learning models such as Support Vector Machines (SVM) and Random Forest (RF), yield high predictive accuracy for overall survival (OS) and progression-free survival (PFS). However, methodological heterogeneity across studies, including variations in imaging protocols, segmentation techniques, and feature selection strategies, underscores the need for standardized AIRI frameworks to enhance comparability and translational applicability (Fig. 8).

By synthesizing existing methodologies and identifying best practices, this study paves the way for a more concise approach to AI-driven radiomics in glioma prognosis. The proposed universal roadmap outlines critical steps from data acquisition to model development, emphasizing automated workflows for segmentation, feature extraction, and AI model training. Implementing these strategies will improve diagnostic consistency, facilitate early intervention and ultimately enhance patient outcomes (Figs. 4, 5, 6, 7, 8).

Future research should focus on refining AI models with multi-center datasets, integrating molecular and clinical biomarkers, and adopting rigorous validation protocols to bridge the gap between research and real-world application. By fostering interdisciplinary collaboration and adhering to good practices in AIRI, the field can move towards standardized, clinically actionable framework that optimally supports oncologists in patient management and decision-making.

This epistemic meta-analysis gives insights as to how up to today, there is not an established methodology to develop and report AIRI models. On the AIRI table (Tables 2, 3) it becomes evident that some studies report only certain performance indicators and not all of them (AUC, accuracy, sensitivity, and specificity). This helps to raise awareness of the current situation so that it can be amended on future research. We consider this article to be conspicuous and therefore is a vector of change to evolve the AIRI field because it helps to evaluate the field on its’ current state and point out its’ aspect to improve to become a more robust field (Figs. 4, 5, 6, 7, 8).

Furthermore, despite there being AIRI roadmaps, they do not go into the fine details of what to do and not to do on model development to get a robust model and comparable outcomes. This paper proposes specific guidelines and a novel roadmap, which we believe is more complete than others, on the section Proposal of a pipeline to develop and implement an AIRI model for future research to follow. Additionally, the roadmap that we propose at the end, includes how to implement the trained model on a clinical setting, which is a breakthrough.