Skip to content
Frontiers
Search

Perspective

Frontiers in Bioinformatics ·

Volume 6 - 2026 | https://doi.org/10.3389/fbinf.2026.1916468

How AI-based techniques for brain tumour diagnosis can address data variability

  • Paola Lecca1*corresponding author
  • Michela Lecca2
  1. 1Faculty of Engineering, Free University of Bozen-Bolzano, Bolzano-Bozen, Italy
  2. 2Technologies of Vision, Fondazione Bruno Kessler, Trento, Italy
Article metrics0 viewsView details
PDFCiteMetrics

Abstract

Data variability in medical imaging of tumors refers to the inconsistencies in Magnetic Resonance Imaging and Computational Tomography scans caused by heterogeneous tumor biology, varying scanner hardware, and subjective human interpretation. This variability creates significant challenges for accurate diagnoses, treatment tracking, and AI model development. Although the variability in medical imaging data for diagnostic and prognostic purposes is a challenge for AI methods applied to other types of tumors as well, it is particularly problematic when the analysis involves gliomas and glioblastomas. Most gliomas are infiltrative and glioblastomas are highly infiltrative tumors with poorly defined margins. They exhibit significant regional variations, including cystic necrosis, localized hemorrhages, and complex microvascular proliferation. Addressing this variability requires a combination of algorithmic innovation, standardized datasets, and a leading-edge computational pipeline including traditional and AI-based data modelling methodologies to ensure diagnostic and prognostic models are robust. In this article, we will identify some AI methods that show promise in the analysis of brain tumor images and that must overcome the challenges posed by data availability. We will also offer some insights into how these challenges can be overcome and propose a possible computational pipeline that utilizes them.

Keywords

1 Introduction

Gliomas are primary tumors of the central nervous system (brain or spinal cord) that originate from glial cells, making up approximately 40% of brain tumors. Brain gliomas are infiltrative and heterogeneous tumors that pose difficulties in early detection and effective treatment management. Artificial Intelligence (AI) has the ability to transform the field of neuro-oncology and can improve glioma detection, imaging segmentation, and non-invasive molecular characterization more effectively than traditional diagnostic methods through deep learning-based radiomics and radiogenomics (Menze et al., 2021; Voigtlaender et al., 2025). However, data variability serves as a significant challenge to the application of AI in neuro-oncology for the treatment of gliomas. Although AI models demonstrate impressive accuracy in controlled experiment - especially regarding tumor segmentation and classification - variability in data sources, imaging protocols, and patient demographics greatly hinders their generalization and clinical use. Furthermore, substantial variability in molecular, histological, and radiological characteristics, results in unpredictable tumor behavior, treatment reactions, and patient results. This drastic variability is seen both among different patients (inter-tumoral) and within the same patient (intra-tumoral), resulting in a recurrence rate exceeding 90% for glioblastomas (see (Ivy Center, 2026; van der Sanden et al., 2016; Jørgensen et al., 2023; Patricoski-Chavez et al., 2025; Sleptsova et al., 2026) for detailed data on recurrence percentage of gliomas and glioblastomas).

Gliomas range from lower-grade tumors to aggressive glioblastomas, they vary massively in their form, cellular makeup, and location, making standard models difficult to generalize (Reddy et al., 2025). Furthermore, a single glioma can contain varying grades of malignancy, necrotic cores, and surrounding edema (Hu et al., 2020), and tumor growth distorts surrounding brain structures and ventricles, warping the spatial landmarks that automated algorithms rely on. Five ways data variability impairs AI-based glioma diagnosis include the following. (i) Imaging variability: differences in scanners, field strengths, sequences, and preprocessing across institutions cause AI models to perform poorly on new datasets. (ii) Limited datasets: reliance on small, single-center data fails to capture patient diversity, leading to model overfitting. (iii) Temporal contamination: older data lack modern molecular profiling, tainting datasets with obsolete information. (iv) Tumor heterogeneity: high biological, cellular, and genetic variation within and between gliomas makes tumor behavior difficult to predict reliably. (v) Radiomics challenges: high-dimensional imaging features are highly sensitive to variations in segmentation strategies, spatial dimensions, and distinguishing edema from true tumor infiltration. In Supplementary Material, Section 1 explores these aspects in greater depth. In turn, data variability affects the complexity of a deep neural network by increasing the required model capacity, training time, and risk of overfitting, finally contributing to the opaqueness of AI algorithms (more details in Section 2 of Supplementary Material).

First, in this perspective article, based on an in-depth analysis of recent literature on the subject, we wish to highlight the methodologies that can reduce the variability of the input data for the most advanced AI algorithms, whilst also explaining the mechanisms through which data variability can undermine the promises of AI, especially deep neural networks. Secondly, building on the understanding of these mechanisms, we provide a computational pipeline. This pipeline embodies a conceptual framework designed precisely for the translation of raw data characterised by high variability and heterogeneity into predictive models the diagnosis and prognosis of brain tumors.

2 Data variability and AI-based approaches facing its challenges

Addressing data variability in gliomas with AI necessitates a comprehensive approach that considers both the natural diversity of the tumor and the differing technical aspects of imaging and molecular data among various institutions (AI models often fail to generalize across hospitals due to protocol-specific biases). Essential strategies encompass the following actions.

  • Applying uniform data preprocessing and annotation standards. Variability in imaging techniques (scanners, magnetic field strengths, reconstruction algorithms) leads to batch effects, i.e., non-biological, technical differences in medical images. These effects can lead to incorrect, biased results when training artificial intelligence models or analyzing multi-site data. Batch effects are particularly problematic for radiomics and deep learning, as they introduce variations in intensity, contrast, and texture that are not related to the patient’s biology (Hu et al., 2023; Stamoulou et al., 2022). Practical implementation of data preprocessing to minimize data variability consists of (i) normalizing data through strict pre-processing methods, such as logarithmic transformation, quantile normalization, and z-score scaling, (ii) employing tools to unify multi-platform datasets (for instance, ComBat, Harmony, or Limma (Tran et al., 2020; Chazarra-Gil et al., 2021; Lee et al., 2023; D’Anna et al., 2024; Patikas et al., 2026)), and (iii) implementing coordinated annotation protocols such as tackling noise in expert assessments by developing standardized, and multi-observer agreement directives for tumor segmentation to minimize human-generated bias.
  • Leveraging AI frameworks such as federated learning (FL) (Li et al., 2026b; Saeed et al., 2025; Moshawrab et al., 2023) and generative models to improve model generalizability. FL is specifically designed to address data variability, diversity, and heterogeneity problems in machine learning, particularly when data is scattered across multiple, decentralized locations. While traditional machine learning requires collecting data in one place - which often fails due to privacy constraints or data silos - federated learning is a decentralized machine learning approach where AI models are trained across multiple devices or edge servers (like smartphones or IoT devices) without sharing raw data. Instead of centralizing data, devices train a local model and send only parameter updates to a central server to improve a shared global model, enhancing privacy and reducing communication overhead. FL would enable AI models to learn from decentralized data across various institutions without exchanging raw patient information, maintaining privacy while benefiting from diverse, varied datasets.
  • Integrating radiomics features from MRI with molecular, histological, and clinical data to create a comprehensive biological landscape of the tumor. Using advanced AI techniques (e.g., deep learning feature fusion (Zhang et al., 2024)) to combine imaging features with omics data (genomics, transcriptomics) improves prediction of survival and treatment response.
  • Using AI techniques to explore peritumoral areas. These zones in glioblastomas are not merely uniform edema zones; they contain infiltrating tumour cells and varied immune responses often overlooked by human specialists (Zhang et al., 2023; Elmahdy and Sebro, 2023; Azemi and Di Ieva, 2025; Liu H. et al., 2025; Linton-Reid et al., 2026). Variability in data within the peritumoral region is a primary concern in advanced oncology and medical imaging. Since this area serves as the battleground between the tumor and the host’s immune system, changes in its biological and imaging characteristics offer essential information regarding disease severity and treatment effectiveness (Xu et al., 2024; Liu Y. et al., 2025; Zhao Q. et al., 2025). The peritumoral region is irregular. It shows organized spatial gradients of biological activity that vary with increasing distance from the tumor edge (Triantafyllou et al., 2026). in Table 1, we identify four core AI techniques that show promise for the analysis of brain tumours, but are particularly hampered by data variability and by the characteristics of the peri-tumoural area, which is so important for prognostic purposes.
  • Explainable AI (XAI) glioma detection mitigates data variability in medical imaging by contributing transparency in deep learning decisions (Muhammad and Bendechache, 2024; Marchi, 2026). XAI hold aims to transform unclear predictions into understandable visual information, fostering clinical trust, facilitating subgroup-specific audits to reveal concealed algorithmic bias, and averting misdiagnoses that may occur when AI models aim to generalize across diverse clinical environments (Ray, 2026). In other words, XAI helps us understand how the variability of data processed using AI techniques affects the results in two ways: by extreme context-specificity and undesirable generalization. See Supplementary Material, Section 3 for further details.

TABLE 1. Data variability in neuro-oncology significantly hampers AI techniques. Here we list the causes of data variability that have the greatest impact on AI performance in the various areas of neuro-oncology in which it is applied.

AI techniquesCauses of data variability
AI-assisted radiomicsRadiomics examination of the imaging area adjacent to a tumor can benefit from the usage of earning to identify microstructural, vascular, and immunological alterations in the tissue stroma or edema, greatly improving the precision of cancer staging, prognosis, and treatment response forecasts (Dou et al., 2018; Li et al., 2026a). Due to traditional imaging primarily emphasizing active, contrast-positive centers, the non-enhancing, invasive borders tend to be inadequately depicted. Nonetheless, the recent literature reports about virtual biopsy, a combination of sophisticated radiomics and machine learning able to categorize molecular subtypes without the need for invasive biopsies (Holtkamp et al., 2025; Liu et al., 2021; Murray et al., 2022)
Convolutional neural networks (CNNs)CNNs are utilized to examine the spatial and temporal variability of the peritumoral area, frequently surpassing conventional techniques. CNNs can effortlessly detect hierarchical, pixel-level characteristics in 3D and 4D medical imaging at multiple scales, identifying subtle microstructural and perfusion variations that manual radiomics and traditional texture analysis may miss (Wang et al., 2024; Li et al., 2026a)
Multimodal fusion models (MFMs)These models combine various imaging information to illustrate tumor infiltration, particularly in gliomas. They integrate the detailed anatomy provided by standard MRI (T1, T2, FLAIR) with functional, microstructural, and metabolic information from advanced methods (ADC, DTI, perfusion) to define tumor borders beyond typical contrast enhancement. CNNs serve as the core of MFMs, capturing detailed local features from medical images to enhance the overall context offered, for instance, by multi-omics data (Alleman et al., 2023; Yi et al., 2025; Saueressig et al., 2026). MFMs are promising tools (Liang et al., 2025), but they are not without challenges that still need to be overcome. Specifically, we identify two major challenges as follows. Multimodal architectures often require multi-branch deep learning networks with independent backbones (e.g., CNNs for imaging). Scaling these simultaneously dramatically increases parameter counts, tensor sizes, and memory footprints, making real-time processing or rapid training on standard hardware difficult without significant optimization (Al-Zoghby et al., 2025). Furthermore, harmonizing different modalities (such as clinical text, genomic sequences, and high-resolution medical images) is non-trivial. Modalities exhibit heterogeneous structures, variable sampling rates, and different semantic resolutions. Bridging this gap requires advanced alignment methods (e.g., cross-modal attention (Brindha and Kavitha, 2026; Driver and Spence, 1998)) to weight data properly without losing modality-specific characteristics
Machine learning classifiersAlgorithms such as support vector machines (SVM) (Wang, 2022), Random forests (Breiman, 2001), XGBoost (Chen and Guestrin, 2016), and CatBoost (Prokhorenkova et al., 2018) are commonly used to train on these features to predict recurrence. Through the examination of the broad array of features derived from radiomics, machine-learning methods can improve tumor diagnosis, evaluate treatment response, and forecast patient outcomes. More specifically, ML classifiers examine the peritumoral region of gliomas and glioblastomas to distinguish between pure vasogenic edema and infiltrative tumor cells, forecast recurrence, and non-invasively map molecular signatures (Tabassum et al., 2023; Michalska-Foryszewska et al., 2024)

2.1 Data variability as a reason of lack of feature reproducibility

The variability of radiomic data also affects the reproducibility of radiomic features. Unstandardized image preprocessing (e.g., interpolation, intensity discretization) and differing computational algorithms frequently lead to discordant feature values on identical datasets. Reproducibility can be established in two stages: imaging reproducibility (i.e., test-retest stability of a scan) and statistical reproducibility (i.e., robustness of features to segmentation and feature extraction software) (Ibrahim et al., 2021). Feature reduction and compliance to Image Biomarker Standardization Initiative (IBSI) are proposed to ensure mathematical harmonization, software interoperability, and cross-institutional reproducibility of MRI-derived tumor phenotypes, but they are not without their drawbacks, as highlighted in Supplementary Table S1 in Supplementary Material.

It should also be noted that reproducibility and standardization are not independent of transparency, as discussed in Supplementary Material, Section 4).

3 An operative pipeline to mitigate the effects of data variability

The strategies we have listed in points a) – e) in the previous section as possible actions to be taken to ensure that the impact of data variability on AI methodologies is mitigated and that, at the same time, AI can manage this variability without compromising its modelling and predictive power, do not, when used in isolation, solve the problem of data variability, but merely address specific, context-dependent aspects of it. As a way forward, we propose a possible framework in which each of the listed strategies works in concert with the others, as illustrated in Figure 1. The framework is organised into three levels: the first level, in which federated learning and data pre-processing procedures operate in parallel to standardise the data. There is then a second level that processes the outputs obtained in the first layer and involves data integration which integrates imaging characteristics with clinical data, genomics, and multi-omics to enhance precision medicine, surpassing single-source models in forecasting disease advancement and treatment responses (Farina et al., 2023; Marouf et al., 2025). AI techniques (radiomics analysis, CNNs, Multimodal fusion models and machine learning classifiers), situated at the third level of the data processing and modelling framework, are applied to the integrated data. We anticipate that, in the immediate and near future, as far as glioma and glioblastoma are concerned, these techniques will need to be tailored to the peri-tumoural area not only because this zone is a primary site of microscopic tumor cell invasion and the source of the majority of recurrences following surgery (Trevisi and Mangiola, 2023; Ballestín et al., 2024), but also because its spatial heterogeneity is not well studied (Wang et al., 2024), and AI techniques should be used, tested and empowered to explore, identify and prognostically classify these heterogeneities. At the third level we also place the XAI post hoc explainers, which employ methods such as feature attribution models (e.g., SHAP, lime) to emphasize crucial microenvironment features on the brain image to aid radiologists in the diagnosis. XAI clarifies imaging in three ways (Lu et al., 2025): (i) by highlighting which texture or shape metrics drive a risk score: (ii) by pinpointing exact zones in the tissue surrounding a tumor influencing an stereotactic body radiotherapy outcome prediction, and (iii) by replacing hidden computations with clear visual evidence matching medical knowledge. By pinpointing the specific quantitative characteristics in the tissue around a tumor that influence an AI’s decision, doctors can validate and trust non-invasive biomarkers for early diagnosis and treatment strategies. Post-hoc explainers can be utilized, as they offer numerical values indicating the precise contribution of a particular peritumoral texture to the model’s prediction, thus enhancing the transparency and clinical reliability of the AI’s reasoning.

Illustrative diagram of a framework for the preprocessing of biological and medical data and for their processing with artificial intelligence techniques specifically designed to support the physician in the evaluation of brain tumor lesions such as gliomas and glioblastomas. Starting with clinical and biological data as input, and using methodologies and objectives that we classify into three levels (federated learning and data standardisation (Level 1), data integration (Level 2), and finally the application of AI techniques (Level 3)), the computational framework produces an output consisting of the generation of information and the extraction of features to support diagnosis and the identification of prognostic biomarkers. The pathway from Level 1 to Level 3 is not without its obstacles or challenges. Specifically, the data integration process is a particularly delicate and important stage, given that the integration of MRI, and PET scans with genomic and clinical profiles transforms glioma diagnostic analyses into a non-invasive, data-driven process. The greatest challenge that the data integration process is likely to face in the near future is the choice of a suitable integration method that mitigates overfitting (if not prevent) – a risk that is becoming increasingly likely due to the continuous refinement of data integration techniques and the ever-increasing volume of data to be integrated. Explainable AI plays also an important role in the pipeline, as it tackles the “black box ”characteristic of deep learning in neuro-oncology by elucidating which MRI or genetic attributes influence AI models. XAI applied to the peritumoral brain are can help clinicians differentiate between tumor infiltration and vasogenic edema on non-invasive advanced neuroimaging.
Figure 1. Illustrative diagram of a framework for the preprocessing of biological and medical data and for their processing with artificial intelligence techniques specifically designed to support the physician in the evaluation of brain tumor lesions such as gliomas and glioblastomas. Starting with clinical and biological data as input, and using methodologies and objectives that we classify into three levels (federated learning and data standardisation (Level 1), data integration (Level 2), and finally the application of AI techniques (Level 3)), the computational framework produces an output consisting of the generation of information and the extraction of features to support diagnosis and the identification of prognostic biomarkers. The pathway from Level 1 to Level 3 is not without its obstacles or challenges. Specifically, the data integration process is a particularly delicate and important stage, given that the integration of MRI, and PET scans with genomic and clinical profiles transforms glioma diagnostic analyses into a non-invasive, data-driven process. The greatest challenge that the data integration process is likely to face in the near future is the choice of a suitable integration method that mitigates overfitting (if not prevent) – a risk that is becoming increasingly likely due to the continuous refinement of data integration techniques and the ever-increasing volume of data to be integrated. Explainable AI plays also an important role in the pipeline, as it tackles the “black box ”characteristic of deep learning in neuro-oncology by elucidating which MRI or genetic attributes influence AI models. XAI applied to the peritumoral brain are can help clinicians differentiate between tumor infiltration and vasogenic edema on non-invasive advanced neuroimaging.

Level 3 is central to the pipeline shown in Figure 1, as it effectively serves as the foundation for the implementation of non-invasive analyses for diagnostic and prognostic purposes. Choosing an appropriate data fusion method is strategic to forecast particular molecular subtypes (e.g., wild-type versus mutant) and identify microscopic infiltration areas prior to surgical biopsy. The data fusion methods commonly used are known as early fusion, late fusions and intermediate fusion. Early fusion which combines radiomic and genomic data prior to entering them into a diagnostic model. This enables profound cross-modal interactions, and that is a benefit since knowing these interactions means knowing how the data and models influence, inform, and combine with each other, but it presents an increased chance of overfitting (Zhao L. et al., 2025; Marouf et al., 2025). Late fusion treats each data type (e.g., MRI scans vs. genomics) using separate, independent workflows, merging their results solely at the final decision-making stage. Since the models do not interact during the training process, late fusion may miss intricate, cross-modal interactions or cooperative biomarkers that develop in the earlier phases of representation. To address this, researchers frequently investigate intermediate fusion or hybrid architectures that link networks prior to the final decision. Hybrid/intermediate fusion techniques utilize sophisticated deep learning frameworks, including Transformer-3D CNN hybrid models featuring cross-attention. These networks focus on the most crucial genomic markers, such as isocitrate dehydrogenase (IDH), while also examining spatial MRI/PET characteristics. IDH mutations are indeed critical diagnostic and prognostic genomic markers primarily used in the classification of diffuse gliomas (Solomou et al., 2023).

4 Discussion

At the moment of writing, computational pipelines that integrate data variability mitigation methods at different levels of a computational global, strongly data-integrative pipeline and address various aspects of the raw data to be processed are largely absent from the literature. In literature we can find, however, computational pipelines that specifically outline the steps required to apply each of the single approaches we organised into the three levels of Figure 1. Unfortunately, in the majority of cases, these computational frameworks have limited clinical application due to the fact that standardizing biological and imaging metrics is complex. The reason lies in the fact that standardising data is a complex process, and for this very reason, fragmenting this standardisation by confining it to specific purposes and methods not only contributes to increasing the uncontrollability of data variability (Taha et al., 2021), but it also represents a considerable effort that does not pay off in terms of achieving the desired outcome, namely, a diagnosis that is interpretable, verifiable and accurate.

The pipeline we propose is a hypothetical one, a vision, in fact, but compared to current proposals, which are context- and method-specific and yet often remain purely conceptual, it aims to address the problem of mitigating data variability in neuro-oncology on a global scale. It therefore presents itself as a challenge that may be worth tackling, given that the benefits in terms of supporting neuro-oncological diagnosis could be significant. On the contrary, the pipeline we propose in Figure 1 aims to address the problem of mitigating data variability in neuro-oncology on a global scale. It therefore presents itself as a challenge that may be worth tackling, given that the benefits in terms of supporting neuro-oncological diagnosis could be significant. At the present, the state-of-art conceived strategy to approach the issue of data variability is a pipeline-based fusion that sequentially integrates image-level harmonization, robust deep learning architectures, and statistical correction techniques. This multi-stage pipeline provides strong robustness, but combining these methods introduces distinct trade-offs across four main areas, e.g., (i) data distortion and artifact generation (especially if generative AI methods are used), (ii) high computational and data demands, (iii) pipeline complexity and cascading errors, (iv) destruction of biological meaningful variance. See for instance (Ouhmouk et al., 2025; Ho et al., 2026; Khalighi et al., 2024; Bakas et al., 2024; Taha et al., 2021; Kamga et al., 2025; Li and Huang, 2020), as example of literature references highlighting the issues (i) - (iv). Problems (i) and (iv) are particularly undesirable for the following reasons. Generative AI (such as diffusion models and Generative Adversarial Networks (Yazdani et al., 2025)) can introduce artificial, non-existent features (hallucinations or artifacts) into synthetic medical data. While used to augment MRI scans or predict tumor growth, generated outputs carry the risk of presenting false anatomical structures that could lead to misdiagnosis or flawed surgical planning (Xia et al., 2025; Parchur et al., 2025). Concerning the statistical corrections, once the statistical correction factors [like ComBat parameters (Reynolds et al., 2023)] are locked post-training, they cannot dynamically adapt to real-time, single-patient clinical inference without recalculating cohort statistics. Since our pipeline does not involve the use of generative AI nor the implementation of statistical corrections, it aims to reduce, if not eliminate, the risks associated with trade-offs (i) and (iv). Regarding the elimination of statistical adjustments to avoid the risk of eliminating biological relevant variance (trade-off (iv)), we clarify that, indeed, statistical corrections in AI can inadvertently eliminate or obscure biologically relevant variance. While harmonization techniques are vital for mitigating technical artifacts Bakas et al. (2024); Dommalapati et al. (2025); Voigtlaender et al. (2025); Hu et al. (2023); Beer et al. (2020); Nan et al. (2023); Reynolds et al. (2023), over-correction poses a significant risk to biomarker discovery and personalized medicine (Hui et al., 2024).

In conclusion, mitigating the variability of the input data for these methods can therefore be beneficial when analysing such a heterogeneous area. Whilst research into AI methods in radiomics (Level 3 in the pipeline) is making great strides, it is becoming crucial to focus the research more on two delicate stages: data integration and the use of XAI for data processing in peritumoural area. Both data integration and XAI act as bridges between humans and machines, which must be robust and reliable to avoid falling into medical errors with unfortunate outcomes.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fbinf.2026.1916468/full#supplementary-material

Supplementary material Supplementary material at the source

Footnotes

Edited by: Chang Li, St. Jude Children’s Research Hospital, United States

Reviewed by: Rafail Christodoulou, Stanford University, United States

Statements

Conflict of interest

The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

The author PL declared that they were an editorial board member of Frontiers at the time of submission. This had no impact on the peer review process and the final decision.

Funding

The author(s) declared that financial support was not received for this work and/or its publication.

Data availability

The original contributions presented in the study are included in the article/Supplementary Material, further inquiries can be directed to the corresponding author.

Author contributions

PL: Writing – review and editing, Methodology, Writing – original draft, Conceptualization, Visualization. ML: Writing – review and editing, Writing – original draft, Methodology.

Acknowledgements

The authors thanks Mohsen Farsad, Leda Lorenzon, Alessandro Fracchetti, Viktorija Sokolova of the Department of Nuclear Medicine, Provincial Hospital of Bolzano (SABES-ASDAA), Teaching Hospital of the Paracelsus Medical Private University, Bolzano–Bozen, Italy, and Gilbert Spizzo in the Department of Internal Medicine, Oncologic Day Hospital, Hospital of Bressanone (SABES-ASDAA), Teaching Hospital of the Paracelsus Medical Private University, Bressanone–Brixen, Italy, for the helpful data, explanations and discussions on the glioma pathology and the expectations that the medical world has regarding the application of AI methodologies to aid data analysis for diagnostic and prognostic purposes.

References

  1. Al-Zoghby A. M., Ismail Ebada A., Saleh A. S., Abdelhay M., Awad W. A. (2025). A comprehensive review of multimodal deep learning for enhanced medical diagnostics. Comput. Mater. and Continua 84, 4155 – 4193. doi:10.32604/cmc.2025.065571
  2. Alleman K., Knecht E., Huang J., Zhang L., Lam S., DeCuypere M. (2023). Multimodal deep learning-based prognostication in glioma patients: a systematic review. Cancers 15, 545. doi:10.3390/cancers15020545
  3. Azemi G., Di Ieva A. (2025). Radiomic fingerprinting of the peritumoral edema in brain tumors. Cancers 17, 478. doi:10.3390/cancers17030478
  4. Bakas S., Vollmuth P., Galldiks N., Booth T. C., Aerts H. J. W. L., Bi W. L. (2024). Artificial intelligence for response assessment in neuro oncology (ai-rano), part 2: recommendations for standardisation, validation, and good clinical practice. Lancet Oncol. 25, e589 – e601. doi:10.1016/S1470-2045(24)00315-2
  5. Ballestín A., Armocida D., Ribecco V., Seano G. (2024). Peritumoral brain zone in glioblastoma: biological, clinical and mechanical features. Front. Immunol. 15, 1347877. doi:10.3389/fimmu.2024.1347877
  6. Beer J. C., Tustison N. J., Cook P. A., Davatzikos C., Sheline Y. I., Shinohara R. T. (2020). Longitudinal combat: a method for harmonizing longitudinal multi-scanner imaging data. NeuroImage 220, 117129. doi:10.1016/j.neuroimage.2020.117129
  7. Breiman L. (2001). Random forests. Mach. Learn. 45, 5 – 32. doi:10.1023/a:1010933404324
  8. Brindha R., Kavitha A. (2026). Optimizing brain tumor diagnosis: the role of 3d attention convolutional-based cross modal distillation. Multidimensional Syst. Signal Process. 37, 4. doi:10.1007/s11045-025-00900-1
  9. Chazarra-Gil R., van Dongen S., Kiselev V., Hemberg M. (2021). Flexible comparison of batch correction methods for single-cell rna-seq using batchbench. Nucleic Acids Res. 49, e42. doi:10.1093/nar/gkab004
  10. Chen T., Guestrin C. (2016). Xgboost: a scalable tree boosting system. Proc. 22nd ACM SIGKDD Int. Conf. Knowl. Discov. Data Min. 16, 785 – 794. doi:10.1145/2939672.2939785
  11. Dommalapati L., Guenter R., Golivi Y., Ganji S. P., Sandhya T., Nagaraju G. P. (2025). Advances in artificial intelligence for spatial transcriptomics in cancer: special focus on yin yang 1 (yy1) and raf kinase inhibitor protein (rkip). Biochimica Biophysica Acta (BBA) - Rev. Cancer 1880, 189456. doi:10.1016/j.bbcan.2025.189456
  12. Dou T. H., Coroller T. P., van Griethuysen J. J. M., Mak R. H., Aerts H. J. W. L. (2018). Peritumoral radiomics features predict distant metastasis in locally advanced nsclc. PLOS ONE 13, e0206108. doi:10.1371/journal.pone.0206108
  13. Driver J., Spence C. (1998). Crossmodal attention. Curr. Opin. Neurobiol. 8, 245 – 253. doi:10.1016/s0959-4388(98)80147-5
  14. D’Anna A., Stella G., Gueli A. M., Marino C., Pulvirenti A. (2024). Mitigating interobserver variability in radiomics with combat: a feasibility study. J. Imaging 10, 270. doi:10.3390/jimaging10110270
  15. Elmahdy M., Sebro R. (2023). Radiomics analysis in medical imaging research. J. Med. Radiat. Sci. 70, 3 – 7. doi:10.1002/jmrs.662
  16. Farina B., Guerra A. D. R., Bermejo-Peláez D., Miras C. P., Peral A. A., Madueño G. G. (2023). Integration of longitudinal deep-radiomics and clinical data improves the prediction of durable benefits to anti-pd-1/pd-l1 immunotherapy in advanced nsclc patients. J. Transl. Med. 21, 174. doi:10.1186/s12967-023-04004-x
  17. Ho B. C., Kim D., Kumar A., Weiss S., Vossler H., Mormino E. (2026). Evaluation of image-level harmonization methods for multi-center mr neuroimaging. J. Magnetic Reson. Imaging 63, 1257 – 1269. doi:10.1002/jmri.70221
  18. Holtkamp M., Parmar V., Hosch R., Salhöfer L., Styczen H., Li Y. (2025). Ai-guided virtual biopsy: automated differentiation of cerebral gliomas from other benign and malignant mri findings using deep learning. Neuro-Oncology Adv. 7, vdae225. doi:10.1093/noajnl/vdae225
  19. Hu L. S., Hawkins-Daarud A., Wang L., Li J., Swanson K. R. (2020). Imaging of intratumoral heterogeneity in high-grade glioma. Cancer Lett. 477, 97 – 106. doi:10.1016/j.canlet.2020.02.025
  20. Hu F., Chen A. A., Horng H., Bashyam V., Davatzikos C., Alexander-Bloch A. (2023). Image harmonization: a review of statistical and deep learning methods for removing batch effects and evaluation metrics for effective harmonization. NeuroImage 274, 120125. doi:10.1016/j.neuroimage.2023.120125
  21. Hui H. W. H., Kong W., Goh W. W. B. (2024). Thinking points for effective batch correction on biomedical data. Briefings Bioinforma. 25, bbae515. doi:10.1093/bib/bbae515
  22. Ibrahim A., Primakov S., Beuque M., Woodruff H., Halilaj I., Wu G. (2021). Radiomics for precision medicine: current challenges, future prospects, and the proposal of a new framework. Methods 188, 20 – 29. doi:10.1016/j.ymeth.2020.05.022
  23. Ivy Center (2026). Brain Tumor recurrence ivy brain Tumor Center — ivybraintumorcenter.org. Available online at: (Accessed June 01, 2026).
  24. Jørgensen A. C. S., Hill C. S., Sturrock M., Tang W., Karamched S. R., Gorup D. (2023). Data-driven spatio-temporal modelling of glioblastoma. R. Soc. Open Sci. 10, 221444. doi:10.1098/rsos.221444
  25. Kamga I. N., Ellison J., Tran N., Phillips J. J., Molinaro A. M., Li Y. (2025). Img-130. integrating ai-based models from physiologic mri with mr spectra for differentiating treatment-effect from glioma tumor recurrence. Neuro-Oncology 27, v305 – v306. doi:10.1093/neuonc/noaf201.1208
  26. Khalighi S., Reddy K., Midya A., Pandav K. B., Madabhushi A., Abedalthagafi M. (2024). Artificial intelligence in neuro-oncology: advances and challenges in brain tumor diagnosis, prognosis, and precision treatment. Npj Precis. Oncol. 8, 80. doi:10.1038/s41698-024-00575-0
  27. Lee H., Seo S., Won S., Park W.-Y., Choi J. Y., Lee K.-H. (2023). Comparative analysis of batch correction methods for fdg pet/ct using metabolic radiogenomic data of lung cancer patients. Sci. Rep. 13, 18247. doi:10.1038/s41598-023-45296-9
  28. Li X. T., Huang R. Y. (2020). Standardization of imaging methods for machine learning in neuro-oncology. Neuro-Oncology Adv. 2, iv49 – iv55. doi:10.1093/noajnl/vdaa054
  29. Li J., Wei J., Pan R., Ma D., Xie Z., Gao Z. (2026a). Multicenter peritumoral radiomics in oncology: advances, challenges, and future directions. Radiat. Oncol. 21, 62. doi:10.1186/s13014-026-02817-w
  30. Li Z., Dib O., Li S. (2026b). Data-centric federated learning for neuro-oncology: addressing heterogeneity via privacy-preserving generative augmentation. Eng. Appl. Artif. Intell. 173, 114351. doi:10.1016/j.engappai.2026.114351
  31. Liang C., Silva R. F., Adali T., Jiang R., Zhang D., Qi S. (2025). Multimodal data fusion in neuroscience: promises, challenges, and future directions [special issue on accelerating brain discovery through data science and neurotechnology]. IEEE Signal Process. Mag. 42, 8 – 21. doi:10.1109/msp.2025.3604570
  32. Linton-Reid K., Chen M., Martell M., Posma J., Aboagye E. (2026). Radiomics in clinical radiology: advances, challenges, and future directions. Clin. Radiol. 92, 107165. doi:10.1016/j.crad.2025.107165
  33. Liu D., Chen J., Hu X., Yang K., Liu Y., Hu G. (2021). Imaging-genomics in glioblastoma: combining molecular and imaging signatures. Front. Oncol. 11, 699265. doi:10.3389/fonc.2021.699265
  34. Liu H., Zhang H., Zou Q., Yang J. (2025a). Intratumoral and peritumoral radiomics of mri predict pathological differentiation in patients with rectal cancer. Oncol. Lett. 31, 1 – 13. doi:10.3892/ol.2025.15363
  35. Liu Y., Pan Y., Wang Q., Jiang H., Lu N., Chen D. (2025b). Integration of intratumoral/peritumoral radiomics and deep learning for predicting overall survival in non-small cell lung cancer patients: a multicenter study. Front. Oncol. 15, 1669200. doi:10.3389/fonc.2025.1669200
  36. Lu Y., Wang Y., Ding Y., Chen D., He W., Zhong W. (2025). Integrating peritumor and tumor ct radiomics features in predicting local control after sbrt in patients with pulmonary oligometastases. Radiat. Oncol. 20, 129. doi:10.1186/s13014-025-02712-w
  37. Marchi G. (2026). Decoding the “black-box”: explainable artificial intelligence towards trustworthy advancement in respiratory medicine. Breathe 22, 250318. doi:10.1183/20734735.0318-2025
  38. Marouf A. A., Rokne J. G., Alhajj R. (2025). Integrating multi-omics and medical imaging in artificial intelligence-based cancer research: an umbrella review of fusion strategies and applications. Cancers 17, 3638. doi:10.3390/cancers17223638
  39. Menze B., Isensee F., Wiest R., Wiestler B., Maier-Hein K., Reyes M. (2021). Analyzing magnetic resonance imaging data from glioma patients using deep learning. Comput. Med. Imaging Graph. 88, 101828. doi:10.1016/j.compmedimag.2020.101828
  40. Michalska-Foryszewska A., Bujko M., Kwiatkowska-Miernik A., Ziemba K., Sklinda K., Walecki J. (2024). The peritumoral brain zone in glioblastoma: a review of the pretreatment approach. Pol. J. Radiology 89, 480 – 487. doi:10.5114/pjr/192044
  41. Moshawrab M., Adda M., Bouzouane A., Ibrahim H., Raad A. (2023). Reviewing federated machine learning and its use in diseases prediction. Sensors 23, 2112. doi:10.3390/s23042112
  42. Muhammad D., Bendechache M. (2024). Unveiling the black box: a systematic review of explainable artificial intelligence in medical image analysis. Comput. Struct. Biotechnol. J. 24, 542 – 560. doi:10.1016/j.csbj.2024.08.005
  43. Murray J. M., Wiegand B., Hadaschik B., Herrmann K., Kleesiek J. (2022). Virtual biopsy: just an ai software or a medical procedure? J. Nucl. Med. 63, 511 – 513. doi:10.2967/jnumed.121.263749
  44. Nan Y., Xing X., Yang G. (2023). Data Harmonization to Address the Non-biological Variances in Radiomic Studies. Springer International Publishing, 95 – 115. doi:10.1007/978-3-031-48446-9_5
  45. Ouhmouk M., Baichoo S., Abik M. (2025). Challenges in ai-driven multi-omics data analysis for oncology: addressing dimensionality, sparsity, transparency and ethical considerations. Inf. Med. Unlocked 57, 101679. doi:10.1016/j.imu.2025.101679
  46. Parchur A. K., Zarenia M., Gage C., Paulson E. S., Ahunbay E. (2025). Automated hallucination detection for synthetic ct images used in mr-only radiotherapy workflows. Phys. Med. and Biol. 70, 05NT01. doi:10.1088/1361-6560/adb5eb
  47. Patikas N., Yao H., Madhu R., Raychaudhuri S., Hemberg M., Korsunsky I. (2026). Integration of Large, Complex single-cell Datasets with harmony2. doi:10.64898/2026.03.16.711825
  48. Patricoski-Chavez J. A., Nagpal S., Singh R., Warner J. L., Gamsiz Uzun E. D. (2025). A deep learning model to predict glioma recurrence using integrated genomic and clinical data. Commun. Med. 5, 359. doi:10.1038/s43856-025-01083-3
  49. Prokhorenkova L., Gusev G., Vorobev A., Dorogush A. V., Gulin A. (2018). Catboost: unbiased boosting with categorical features. Proc. 32nd Int. Conf. Neural Inf. Process. Syst. 18, 6639 – 6649. doi:10.48550/arXiv.1706.09516
  50. Ray P. P. (2026). A review on explainable artificial intelligence in radiomics: State-of-the-art tools, prospective use cases, challenges and future directions. Eur. J. Radiology Artif. Intell. 5, 100069. doi:10.1016/j.ejrai.2025.100069
  51. Reddy S., Lung T., Muniyappa S., Hadley C., Templeton B., Fritz J. (2025). Radiomics and radiogenomics in differentiating progression, pseudoprogression, and radiation necrosis in gliomas. Biomedicines 13, 1778. doi:10.3390/biomedicines13071778
  52. Reynolds M., Chaudhary T., Eshaghzadeh Torbati M., Tudorascu D. L., Batmanghelich K. Alzheimer's Disease Neuroimaging Initiative (2023). Combat harmonization: empirical bayes versus fully bayes approaches. NeuroImage Clin. 39, 103472. doi:10.1016/j.nicl.2023.103472
  53. Saeed N., Ashour M., Mashaly M. (2025). Comprehensive review of federated learning challenges: a data preparation viewpoint. J. Big Data 12, 153. doi:10.1186/s40537-025-01195-6
  54. Saueressig C., Scholz D., Raffler P., Delbridge C., Wiestler B., Schüffler P. (2026). Multimodal fusion of pathology and radiology foundation models for who 2021 glioma subtyping. Npj Precis. Oncol. 10, 118. doi:10.1038/s41698-026-01366-5
  55. Sleptsova E., Vershinina O., Ivanchenko M., Turubanova V. (2026). From data to decision: integrating bioinformatics into glioma patient stratification and immunotherapy selection. Int. J. Mol. Sci. 27, 667. doi:10.3390/ijms27020667
  56. Solomou G., Finch A., Asghar A., Bardella C. (2023). Mutant idh in gliomas: role in cancer and treatment options. Cancers 15, 2883. doi:10.3390/cancers15112883
  57. Stamoulou E., Spanakis C., Manikis G. C., Karanasiou G., Grigoriadis G., Foukakis T. (2022). Harmonization strategies in multicenter mri-based radiomics. J. Imaging 8, 303. doi:10.3390/jimaging8110303
  58. Tabassum M., Suman A. A., Suero Molina E., Pan E., Di Ieva A., Liu S. (2023). Radiomics and machine learning in brain tumors and their habitat: a systematic review. Cancers 15, 3845. doi:10.3390/cancers15153845
  59. Taha B., Boley D., Sun J., Chen C. (2021). Potential and limitations of radiomics in neuro-oncology. J. Clin. Neurosci. 90, 206 – 211. doi:10.1016/j.jocn.2021.05.015
  60. Tran H. T. N., Ang K. S., Chevrier M., Zhang X., Lee N. Y. S., Goh M. (2020). A benchmark of batch-effect correction methods for single-cell rna sequencing data. Genome Biol. 21, 12. doi:10.1186/s13059-019-1850-9
  61. Trevisi G., Mangiola A. (2023). Current knowledge about the peritumoral microenvironment in glioblastoma. Cancers 15, 5460. doi:10.3390/cancers15225460
  62. Triantafyllou M., Giannettou S., Tzortzakakis A., Karantanas A. H., Klontzas M. E. (2026). Peritumoral radiomics indicate a gradient transition between renal tumours and their surrounding healthy tissue: a proof-of-concept analysis. Eur. J. Radiology 194, 112513. doi:10.1016/j.ejrad.2025.112513
  63. van der Sanden B., Ratel D., Berger F., Wion D. (2016). Glioma recurrence following surgery: peritumoral or perilesional? Front. Neurology 7, 52. doi:10.3389/fneur.2016.00052
  64. Voigtlaender S., Nelson T. A., Karschnia P., Vaios E. J., Kim M. M., Lohmann P. (2025). Value of artificial intelligence in neuro-oncology. Lancet Digital Health 7, 100876. doi:10.1016/j.landig.2025.100876
  65. Wang Q. (2022). “ Support vector machine algorithm in machine learning,” in 2022 IEEE International Conference on Artificial Intelligence and Computer Applications (ICAICA), 750 – 756. doi:10.1109/ICAICA54878.2022.9844516
  66. Wang F., Dong J., Xu Y., Jin J., Xu Y., Yan X. (2024). Turning attention to tumor–host interface and focus on the peritumoral heterogeneity of glioblastoma. Nat. Commun. 15, 10885. doi:10.1038/s41467-024-55243-5
  67. Xia M., Bayerlein R., Chemli Y., Liu X., Ouyang J., Lin M. (2025). On hallucinations in artificial intelligence–generated content for nuclear medicine imaging (the dream report). J. Nucl. Med. 125. 270653:174. doi:10.2967/jnumed.125.270653
  68. Xu R., Wang K., Peng B., Zhou X., Wang C., Lu T. (2024). Evaluating peritumoral and intratumoral radiomics signatures for predicting lymph node metastasis in surgically resectable non-small cell lung cancer. Front. Oncol. 14, 1427743. doi:10.3389/fonc.2024.1427743
  69. Yazdani S., Singh A., Saxena N., Wang Z., Palikhe A., Pan D. (2025). Generative ai in depth: a survey of recent advances, model variants, and real-world applications. J. Big Data 12, 230. doi:10.1186/s40537-025-01247-x
  70. Yi G., Ma W., Yu Z., Bai H., Zhang H., Wang Y. (2025). Gliomas analysis via multimodal mri-deep learning fusion: technical innovations in segmentation, molecular subtyping, and clinical translation pathways. Adv. Med. Educ. Pract. 16, 1989 – 2001. doi:10.2147/amep.s554692
  71. Zhang W., Guo Y., Jin Q. (2023). Radiomics and its feature selection: a review. Symmetry 15, 1834. doi:10.3390/sym15101834
  72. Zhang Z., Zhang D., Yang Y., Liu Y., Zhang J. (2024). Value of radiomics and deep learning feature fusion models based on dce-mri in distinguishing sinonasal squamous cell carcinoma from lymphoma. Front. Oncol. 14, 1489973. doi:10.3389/fonc.2024.1489973
  73. Zhao L., Huang S., Li J. (2025a). “ Multimodal data fusion research review: from traditional methods to unsupervised learning methods,” in Proceedings of the 2025 2nd Symposium on Big Data, Neural Networks, and Deep Learning (ACM), 249 – 254. doi:10.1145/3784013.3784052
  74. Zhao Q., Zhang H., Xing W. (2025b). Integrating peritumoral and intratumoral radiomics with deep learning for preoperative prediction of lymphovascular invasion in invasive breast cancer using dce-mri. Technol. Cancer Res. and Treat. 24, 15330338251374945. doi:10.1177/15330338251374945

Citation

Lecca, P. & Lecca, M. (2026). How AI-based techniques for brain tumour diagnosis can address data variability. Frontiers in Bioinformatics, Volume 6. https://doi.org/
Received
Accepted
Published
Volume
6 - 2026

Copyright

Copyright © 2026 Lecca and Lecca.

License: https://creativecommons.org/licenses/by/4.0/

*Correspondence: Paola Lecca

Article metrics

Views
0
Downloads
0
Citations
0
Shares
0

Related articles

Original ResearchIdentification and validation of biomarkers related to oxidative stress in pediatric sepsis utilizing WGCNA and multiple machine learning algorithmsYumeng Liu, Gang LuoReviewStrategic roadmap for FAIR-compliant transcriptomic workflows: from technical challenges to bioinformatics democratizationVerónica Jiménez Jacinto, Leticia Vega-AlvaradoOriginal ResearchMachine learning-based identification of CXCL13 and CXCL10 as key biomarkers in rheumatoid arthritisSubash M, Sheik Asraf S, Santhanamari M, Nagaraj P, Ayodeji Olalekan Salau, Abdullahi Garba Usman, Mouaz Yousfan