médical IA

Topology meets vision transformers for brain tumor classification: what is 99.1% accuracy on a single MRI dataset worth? (Ahmed 2026, arXiv)

Published on June 7, 2026 · 8 min read

Faisal Ahmed (Embry-Riddle Aeronautical University, Arizona) posted on 30 May 2026 on arXiv a model that fuses a vision transformer with topological data analysis to sort brain MRIs into four classes — glioma, meningioma, pituitary tumor, no tumor. It reports 99.10% accuracy and a 99.98% AUC on the public BRISC2025 benchmark. The result is clean and the methodological idea interesting, but the gain over existing models sits within the noise, evaluation rests on a single dataset whose image-by-image split does not rule out data leakage, and the test set is also used to pick the model — a textbook reminder that a near-perfect score is not a clinically validated tool.

The context

Automatically classifying a brain MRI — is there a tumor, and if so which one? — is among the most heavily worked medical-imaging tasks in AI. For a few years now, vision transformers (ViTs: networks that cut an image into small patches and learn the relationships among all of them, rather than sweeping the image with local filters as the older convolutional networks did) have topped the leaderboards. On the standard public benchmarks, the best models now plateau around 98-99% accuracy.

The author starts from a legitimate intuition: a ViT mostly captures the global context of an image, but not explicitly its shape or structure — how regions are connected, whether cavities are present, how intensities are organized geometrically. Yet a tumor has a morphological signature. The idea is therefore to add to the ViT a description drawn from topological data analysis (TDA), a branch of applied mathematics that quantifies the "shape" of an object, and to see whether this complementary signal improves classification.

The method

The work is an arXiv preprint (2606.00927, posted 30 May 2026, 21 pages, CC BY 4.0 license, not yet peer-reviewed). The task is a four-class classification: glioma, meningioma, pituitary tumor, and no tumor.

The model has two branches. The first is a ViT-Base (16-pixel patches, 224×224 input) pretrained on ImageNet, a large dataset of natural images; for each MRI it outputs a vector of 768 numbers summarizing its content. The second branch computes the persistent homology of the image: in practice, one sweeps pixel intensities from darkest to brightest and records at which thresholds connected components appear and disappear, and at which thresholds "holes" form. This is summarized by Betti curves (the number of components and holes at each threshold), condensed into a 198-number vector, itself reduced to 64 by a small network. The two descriptions — 768 from the ViT, 64 from topology — are simply concatenated (832 numbers) and sent to a final classifier that picks one of the four categories.

The data come from BRISC2025, a public dataset of 6,000 contrast-enhanced T1-weighted brain MRIs, annotated by radiologists. The author uses the official split: 5,000 images for training, 1,000 for testing. A detail that will matter: this split is done image by image, and the paper never reports the number of distinct patients. No data augmentation is used, no cross-validation, and the whole thing was trained on a plain Apple M1 laptop with no dedicated graphics card.

The results

On the 1,000 test images, the fused model reaches 99.10% accuracy, 99.27% precision, 99.15% recall, a 99.21% F1, and a 99.98% AUC (the AUC, area under the ROC curve, measures the ability to separate classes: 100% would be perfect separation; the precision, recall and F1 figures are averages over the four classes). On paper, that is near-perfect.

The problem is what it is compared with. The same ViT alone, without the topological branch, already reaches 98.90% accuracy and 99.97% AUC; the topological features alone, classified by a tree algorithm (XGBoost), reach 98.19%. The other cited references — ResNet50, ResNet101, EfficientNetB2 — sit between 98.1 and 98.4%. In other words, the fusion's own contribution is on the order of +0.2 accuracy points and +0.01 AUC points over a comparator already at 98.9%. No confidence interval, no significance test, no repetition over several splits supports the claim that this gap is real rather than the luck of a single draw.

Clinical translation. Let us be blunt: these numbers translate into nothing clinical. 99.1% on 1,000 images means about nine misclassified images — but those are images, not patients, and the "no tumor" class holds only 140 test images, so a handful of errors there would shift the rates sharply. Above all, sorting an already-acquired, already-framed MRI into four broad categories is not the radiologist's task, which is to detect, localize, grade, measure and decide on management. No screening sensitivity or specificity, no comparison to a clinician, no measure against a care endpoint appears in the study.

What is good

A clean methodological idea, backed by an ablation. Coupling an explicit description of shape (persistent homology) with deep features is a legitimate, well-posed avenue. The most convincing point is not the final score but the ablation: the topological features alone, with no deep network at all, already classify at 98.19%. This shows that the image's topology carries a genuine discriminative signal — an interesting result in itself, independent of the fusion.

Minimal, fully reproducible compute. Training fits on an Apple M1 laptop, with no data augmentation and no dedicated GPU. This is the exact opposite of models that reach their scores only at the cost of compute farms: here nothing depends on massive infrastructure, which makes the result easy to reproduce and to challenge. The data are public and the paper is under the open CC BY 4.0 license.

Partial transparency about limits. The author explicitly acknowledges that evaluation is limited to a single dataset and that validation on external sets, under other acquisition conditions and other clinical contexts, remains necessary before talking about generalization. That is precisely the caution most splashy announcements omit.

What is less good

A saturated comparator and a misleading metric. When the comparator — the ViT alone — is already at 98.9%, there is almost no room left to demonstrate an improvement. A 0.2-point gain on a single split, with no confidence interval and no statistical test, is within the noise: one cannot tell a genuine advance from a sampling fluctuation. On a benchmark this saturated, the right reflex is to repeat the experiment over several splits and report a spread; there is nothing of the sort here.

Data leakage not ruled out, and a test set that doubles as validation. The split is done image by image and the number of patients is never given. Yet MRI datasets often contain several slices from the same patient: if some land in training and others in testing, the model recognizes the patient rather than the pathology — that is data leakage, and it artificially inflates the scores. The paper does not control for this. Worse, the test set is also used for early stopping (the best model is kept based on its test performance): the 99.1% is thus measured on data that took part in selecting the model, not on a truly held-out set.

No external validation, no clinical translation, code not shared. Everything rests on a single dataset, a single MRI sequence (contrast-enhanced T1), a single split. We do not know what the model would do on another scanner, another hospital, another population — the classic population bias, never tested here. The task itself stays academic: four broad categories on pre-selected images, far from the real flow of a radiology department. Finally, despite the mention of "prepared" code, no GitHub or Hugging Face repository is provided, which limits verifiability.

What it changes

For the research community. The most useful signal is the topological ablation: persistent homology alone captures most of the information on this dataset. That invites testing topology + deep learning fusion where it could really matter — on non-saturated datasets, with a patient-level split and external validation — rather than scraping tenths of a point on an already-solved benchmark. We discussed the same logic for radiograph foundation models in our decryption of SkelEx.

For clinicians. Nothing changes today. 99% on a clean benchmark is not real-world performance, and the clinically relevant question is not "glioma, meningioma, pituitary or nothing?" on a chosen slice. The gap between a leaderboard score and a usable tool remains wide open.

For patients and the public. This is a good moment to practice skepticism: an announcement like "an AI classifies brain tumors with 99% accuracy" almost always describes a score on a carefully curated test set, not what would happen on your MRI. The number is not false; it simply does not mean what one thinks it means. We gave the same warning about language-model answer formats in radiology in our decryption on GPT-4 and radiology.