A new study published in BMC Medical Imaging offers a glimpse of how that darkness might be lifted before a single incision is made. A team of researchers led by Fang Zhang and Guilian Pan of the Fifth Affiliated Hospital of Zunyi Medical University in Zhuhai, China, developed and validated a machine learning model that combines radiomics and deep learning to predict brain invasion from preoperative magnetic resonance imaging. The work, published on 6 October 2026, draws on a retrospective cohort of 438 patients with pathologically confirmed meningiomas treated at two hospitals, and its central message is both promising and carefully hedged: routine MRI scans, interpreted by algorithms rather than by the naked eye, carry measurable signals of brain invasion that reach moderate predictive strength across independent patient groups.

The technical foundation of the study rests on two complementary ways of extracting information from medical images. Radiomics is the older and more established of the two. It converts the pixels of a medical image into hundreds or thousands of quantitative features that describe the texture, shape, and intensity distribution of a tumor. Features derived from matrices such as the Gray Level Co-occurrence Matrix and the Gray Level Run Length Matrix capture how frequently particular combinations of pixel intensities occur in neighboring regions, while Laplacian of Gaussian and Local Binary Pattern filters highlight fine-scale and multi-scale textural patterns that human readers cannot consciously perceive. In this study, the researchers extracted such features from three standard MRI sequences: T2-weighted imaging, diffusion-weighted imaging, and contrast-enhanced T1-weighted imaging, all of which are acquired routinely in the workup of suspected meningiomas.

Deep learning takes a fundamentally different approach. Rather than relying on hand-crafted descriptors, convolutional neural networks learn their own internal representations directly from the raw image data. The team evaluated several well-known architectures, including VGG11, ResNet-50 in its two-dimensional form, DenseNet-121, and a three-dimensional ResNet-152, applying them to the tumor regions delineated on the MRI scans. Through a technique called global average pooling, the deep networks distilled the spatial patterns they had learned into compact feature vectors that could be fed into a downstream classifier. The researchers then fused these deep features with the radiomic features, creating what they call a DL-radiomics, or DLR, model, which was ultimately built on the random forest algorithm, a robust ensemble method that aggregates the votes of many decision trees to produce a prediction.

Before any model could be trained, the researchers confronted one of the central challenges of medical machine learning: the sheer dimensionality of the feature space relative to the number of patients. Extracting features from three MRI sequences across multiple filter types and deep architectures yields an enormous pool of candidate variables, many of which are redundant, noisy, or correlated with one another. To prune this space, the team applied a three-stage dimensionality reduction pipeline. Univariate logistic regression screened out features with no individual association with brain invasion, correlation analysis removed redundant variables that carried overlapping information, and the Boruta algorithm, an iterative feature selection method built on random forests, identified the subset of features that consistently contributed predictive signal beyond what chance would allow. Feature stability was also assessed using the intraclass correlation coefficient, and feature definitions followed the Image Biomarker Standardization Initiative, an international effort to ensure that radiomic features mean the same thing across studies and institutions.

The results of this pipeline were evaluated across seven predictive models built from different combinations of clinical variables, radiomic features, and deep learning features. Performance was quantified using the area under the receiver operating characteristic curve, or AUC, a standard metric that ranges from 0.5, equivalent to a coin flip, to 1.0, representing perfect discrimination. The DLR model emerged as the strongest performer, achieving an AUC of 0.805 in the training set, 0.795 in the internal validation set, and 0.793 in the external validation set drawn from the second hospital. The narrow spread between these three figures is itself noteworthy, because many machine learning models in medicine perform impressively on the data they were trained on but collapse when confronted with patients from a different institution, where scanner vendors, imaging protocols, and patient demographics may differ. A drop of only about one percentage point from training to external validation suggests the model captured genuine biological signal rather than institutional idiosyncrasies.

Beyond discrimination, the researchers examined whether the model would actually be useful in clinical practice using decision curve analysis, a technique that weighs the benefits of true predictions against the harms of false ones across a range of decision thresholds. The analysis showed a positive net benefit compared with default strategies, such as treating all patients as having brain invasion or treating none as such, but only within a limited threshold range. This nuance matters. Decision curves translate abstract statistical performance into the language of clinical consequences, and the finding indicates that the model could add value for some, but not all, plausible clinical scenarios. The authors also reported that calibration, the agreement between predicted probabilities and observed outcomes, was weaker in the external validation set than in the training set, a common and important limitation that they acknowledged explicitly.

The clinical stakes of preoperative brain invasion prediction are considerable. Under current WHO classification standards, brain invasion upgrades a meningioma to grade 2, which typically triggers more extensive surgical resection, closer radiological follow-up, and consideration of adjuvant radiotherapy in certain cases. If a neurosurgeon knew before the operation that a tumor was likely to invade the brain, the surgical plan could be tailored accordingly, perhaps with wider margins or a lower threshold for postoperative intervention. Conversely, a confident prediction of non-invasion might spare patients with benign-appearing tumors from unnecessarily aggressive procedures. The interpretability analyses included in the study, using SHAP-based explanations for the radiomics and ResNet2D models, offer a window into which image features drove the predictions, an increasingly important requirement for algorithms that hope to earn the trust of clinicians.

Still, the authors are careful to frame their conclusions within the limits of the evidence. This was a retrospective study, meaning the model was applied to scans and outcomes that had already been collected, a design that cannot fully exclude the biases that creep into real-world prospective use. The moderate discrimination, with AUCs hovering near 0.8, is respectable but falls short of the near-perfect performance that would justify standalone clinical decision-making. The weaker external calibration signals that predicted probabilities may need to be adjusted before being trusted at face value in new settings. The authors state plainly that prospective validation and clinical impact evaluation are required, meaning the model must be tested on patients as they are diagnosed, in real time, and shown to actually improve outcomes or decision quality rather than merely to correlate with them.

Even with those caveats, the study adds a meaningful piece to a rapidly growing body of work on AI-assisted neuroradiology. It demonstrates that the information needed to predict an aggressive biological behavior is already embedded in images that hospitals acquire every day, waiting to be decoded by the right combination of feature engineering and machine learning. It also exemplifies the methodological discipline that such claims demand: multi-institutional data, standardized feature definitions, rigorous dimensionality reduction, external validation, and honest reporting of limitations. As radiomics and deep learning models mature from retrospective curiosities into prospective tools, studies of this kind chart the cautious, evidence-driven path by which artificial intelligence may one day stand beside the neurosurgeon, reading the same scans and quietly flagging the tumors that deserve the greatest vigilance.

Subject of Research: Preoperative prediction of brain invasion in meningiomas using MRI-based radiomics and deep learning

Article Title: MRI-based radiomics and deep learning model for preoperative prediction of brain invasion in meningiomas

Article References: Zhang, F., Pan, G., Gui, Y., Ren, J., Su, L., Liu, M., & Zhang, J. (2026). MRI-based radiomics and deep learning model for preoperative prediction of brain invasion in meningiomas. BMC Medical Imaging. https://doi.org/10.1186/s12880-026-02867-2

Image Credits: AI Generated

DOI: 10.1186/s12880-026-02867-2

Keywords: meningioma, brain invasion, radiomics, deep learning, MRI, machine learning, random forest, convolutional neural network, neuroradiology, predictive medicine, Boruta algorithm, WHO grading

The answer, according to the data, appears to be no. Meningiomas classified as atypical—or World Health Organization grade 2—purely on the basis of brain invasion alone showed substantially better progression-free survival than atypical meningiomas that earned that designation through other histologic features. The distinction is not an academic curiosity. It touches on decisions about follow-up imaging schedules, adjuvant radiation therapy, and the level of anxiety patients carry home after leaving the operating room.

Meningiomas are among the most common primary brain tumors in adults, and most behave in an indolent fashion, growing slowly or not at all and often requiring nothing more than observation. The WHO classification system, which pathologists use to grade these tumors, underwent a pivotal change in 2016 when brain invasion—the microscopic presence of tumor cells infiltrating the brain parenchyma—was elevated to a major criterion that could single-handedly justify a grade 2 designation. Under the current system, a tumor with completely benign-appearing cells can still be labeled atypical if the pathologist identifies tumor nests threading into neural tissue. At the same time, grade 2 can also be assigned on the basis of mitotic activity, meaning the tumor is dividing more rapidly than normal, or on the basis of specific cellular features such as sheeting architecture, prominent nucleoli, small cell change, hypercellularity, or spontaneous necrosis.

The trouble, the study’s authors argue, is that brain invasion may not be the same biological signal as these other features. Tumors that grow large can, through sheer mechanical force and mass effect, press into and ultimately indent the brain, potentially allowing tumor cells to become entrapped along the irregular surface of the cortex. In that scenario, brain invasion could be a consequence of a tumor’s size and location rather than evidence of an intrinsically aggressive biology. A tumor that has simply been pushed against the brain, in other words, is not necessarily a tumor programmed to grow back.

To test that hypothesis, the researchers conducted a single-institution retrospective cohort study of patients who underwent meningioma resection between 2003 and 2023. The team meticulously reviewed pathology reports and regraded every tumor according to current WHO criteria, ensuring the analysis reflected contemporary diagnostic standards rather than shifting historical definitions. From this process, they isolated patients whose tumors fell into a category they designated BIOB—”Brain Invasion Otherwise Benign”—meaning these tumors were graded 2 solely because of brain invasion, with none of the other atypical features present. These patients were compared with the “other criteria” group, whose grade 2 designation rested on histologic characteristics other than invasion alone.

The cohort ultimately comprised 131 patients with a mean age of 59.2 years. The median preoperative tumor volume was a substantial 34.3 cubic centimeters, which is notable in itself—these were, on average, sizable tumors before surgery. Gross total resection, the neurosurgeon’s gold standard in which all visible tumor is removed, was achieved in 53 percent of cases. Fourteen percent of patients received adjuvant radiation therapy after their operations. Brain invasion was documented in 60 percent of the cohort, and of those, 45 patients—about a third of the total—formed the BIOB group. Mean radiographic follow-up stretched to 74 months, more than six years, giving the data genuine long-term weight.

The results were striking. Across the entire cohort, tumor progression occurred in 31 percent of patients, at an average of roughly 40 months after resection. But when the researchers split the cohort, the picture diverged sharply. Only 17.8 percent of BIOB patients—eight of forty-five—experienced progression, compared with 37.2 percent of patients in the other criteria group—thirty-two of eighty-six. In other words, patients whose tumors were graded upward solely because of brain invasion were less than half as likely to see their tumors return.

The statistical analysis reinforced what the raw numbers suggested. On univariate Cox regression, membership in the BIOB group was significantly associated with longer time to recurrence, with a hazard ratio of 0.44 and a p-value of 0.036. Gross total resection was also strongly protective, carrying a hazard ratio of 0.29. Critically, when the researchers adjusted for other variables in multivariate analysis, the BIOB designation remained a favorable independent predictor, confirming that the effect was not explained away by tumor size, extent of resection, patient age, or radiation treatment. Notably, the two groups were well matched at baseline: demographic, radiologic, and treatment characteristics did not differ significantly between them, with all comparisons yielding p-values above 0.3. The only hint of a difference was a trend toward younger age in the BIOB group—55.7 years versus 61.1 years—which fell just short of statistical significance.

The implications reach beyond meningioma nomenclature. In current practice, a grade 2 designation typically triggers more intensive surveillance MRI schedules, earlier consideration of radiotherapy for residual or recurrent disease, and a more cautious posture overall. Patients diagnosed with atypical meningiomas face the psychological burden of knowing their tumor is labeled aggressive. If a meaningful subset of grade 2 meningiomas—those defined by invasion alone—biologically resemble their benign counterparts, then these patients may be over-treated, over-monitored, and over-worried. Conversely, the study offers reassurance that the genuinely ominous features are the proliferative and cytologic hallmarks: rapid cell division, necrosis, and the constellation of atypical cellular changes.

The authors were careful to frame their findings within the study’s limitations. As a retrospective, single-institution analysis, it is vulnerable to selection bias and to the inherent variability in how pathologists sample and interpret brain invasion. Identifying invasion requires adequate tissue sampling at the tumor–brain interface, and under-sampling can miss invasion just as easily as over-enthusiastic interpretation can overcall it. The distinction between a tumor that mechanically compresses the brain and one that truly invades it can be genuinely difficult at the microscope, and interobserver variability among neuropathologists is well documented. These caveats have long been the central argument of skeptics who questioned whether brain invasion belonged in the grading system at all.

Yet the study’s methodological choices were designed to blunt those concerns. By regrading all tumors under current WHO criteria and by strictly separating tumors whose only grade-elevating feature was invasion from tumors with other atypical features, the researchers created a natural experiment: if invasion alone carried the same prognostic weight as the other criteria, progression rates in the BIOB group should have matched the other criteria group. Instead, they diverged dramatically, despite nearly identical baseline characteristics.

The mortality data add further texture to the story. Fifteen percent of the cohort died during follow-up, and among patients whose tumors did recur, a range of salvage treatments were deployed, including stereotactic radiosurgery in 18 percent, repeat surgical resection in 6 percent, fractionated stereotactic radiotherapy in 3 percent, and intensity-modulated radiation therapy in 2 percent. The mean tumor volume at the time of progression was 17.4 cubic centimeters, underscoring that even recurrent disease in this population was often detected and treated at manageable sizes—a point in favor of the surveillance protocols currently in place.

What emerges from this work is a compelling case for refining, rather than abandoning, the current grading framework. Brain invasion may still deserve attention as a descriptive finding—something a neuropathologist notes, a surgeon knows about, and a surveillance plan accounts for—without automatically consigning a patient to the full prognostic weight of an atypical diagnosis. Future studies incorporating molecular markers, which have increasingly demonstrated power to stratify meningioma risk beyond histology, may help clarify whether brain invasion interacts with tumor biology in ways the microscope alone cannot reveal. Until then, patients whose meningiomas are graded 2 on the basis of brain invasion alone now have data-driven grounds for a more optimistic conversation with their physicians.

For the broader neuro-oncology community, the study is a reminder that classification systems, however authoritative, are hypotheses about biology—and hypotheses must be tested against outcomes. When the outcomes part ways with the label, it is the label that should evolve.