A centroid-based approach aims to help AI count tiny brain tumors

A brain MRI does not show a tumor as a ready-made object with a neat border. Software must infer which voxels belong to abnormal tissue—and, when there are several lesions, where each one is. Those tasks are related, but not identical: outlining a region is segmentation; identifying separate lesions and estimating their number is instance-level detection and counting.

The authors of MicroBT argue that small lesions are particularly easy for standard training methods to overlook. Many segmentation systems use overlap-based scores, which reward agreement between predicted and reference regions. Because a large tumor covers many more voxels than a tiny one, its errors can weigh more heavily in training. A model might therefore perform well overall while missing small lesions or merging nearby ones. This matters especially for brain metastases, where lesion locations and counts can inform treatment planning. The paper explains the clinical motivation; it does not test whether using MicroBT improves patient outcomes.

Let the lesion’s centre guide the search

MicroBT takes four MRI sequences as input—T1, contrast-enhanced T1, T2 and FLAIR—and processes them with a three-dimensional image encoder. Its central idea is to predict a heatmap of likely lesion centres. The authors treat a centre, or centroid, as a useful anchor regardless of lesion size: a tiny lesion and a large one both have a location that can guide the model.

That predicted centre map feeds two branches. One produces a dense segmentation, marking tumor-containing voxels; the other proposes individual lesion locations and bounding boxes. In this design, the same spatial cue helps the model draw each lesion and count it as a separate instance, rather than asking an independent detector to search the whole volume without that guidance.

The training recipe adds two components. A volume-calibrated loss gives smaller lesions more influence in the localization part of training, aiming to prevent their signals from being swamped by larger tumors. A size-aware curriculum first trains the system without that extra weighting, then introduces it while reducing how quickly the shared image encoder changes. The authors’ rationale is that the model should first learn broad image features before focusing more strongly on fine detail.

Tests on four public datasets

The study evaluates MicroBT on four public BraTS challenge datasets covering metastases, meningiomas and gliomas. The paper says the datasets were prepared using four MRI modalities and that the analysis retained 1,341 cases containing at least one lesion no larger than 3 millimetres. The authors report a random, case-level training and test split within each dataset, with no cases shared between the two sets.

They compare MicroBT with nine segmentation approaches and a set of counting methods. Segmentation is assessed with Dice, which measures overlap, and HD95, a measure of boundary distance. Counting and localization are assessed with mean absolute percentage error (MAPE) and average precision at an intersection-over-union threshold of 0.5 (AP50). In plain terms, these ask how far predicted counts are from reference counts and how well predicted lesion locations match annotated ones.

Across the three tumor types, the paper reports an average Dice score of 82.12% and average HD95 of 13.64 mm for MicroBT. For comparison, nnU-Net—the strongest segmentation baseline in the reported overall results—has an average Dice of 80.51% and HD95 of 17.01 mm. For counting, MicroBT’s reported average MAPE is 0.463 and AP50 is 71.35%; the corresponding figures for nnMamba, the strongest counting baseline by average MAPE, are 0.549 and 63.38%.

The authors also examine lesions at or below 3 mm. In that size group, MicroBT has the best average MAPE among the compared methods, at 0.764, versus 0.920 for nnU-Net and 0.897 for nnMamba. These are benchmark comparisons, not guarantees that every small lesion will be found. The paper’s own reported detection analysis suggests the smallest lesions remain difficult: MicroBT’s sensitivity is about 68% on the ≤3 mm subset at low false-positive rates, according to the authors’ summary of the FROC results.

What the evidence can—and cannot—show

The paper includes component-removal experiments, which support the authors’ view that centroid guidance, size weighting and the training curriculum each contribute to performance. But the evidence is still limited to retrospective tests on the specified datasets and splits. The supplied paper text does not describe a prospective clinical evaluation or demonstrate performance across routine hospital workflows, scanners and patient populations beyond these benchmarks.

There is also a reporting detail worth treating cautiously: the abstract summarizes metastasis gains as roughly 7% better micro-lesion localization precision and 26% lower counting error, while the main results give different comparison metrics and relative changes for particular baselines. These figures may refer to different measures or comparisons; they should not be collapsed into a single general claim.

MicroBT is described in an arXiv preprint, not as a peer-reviewed clinical validation. Its results are claims by the authors and have not, on the evidence supplied here, been independently confirmed. The work offers a specific strategy for a real technical problem—giving tiny lesions a stronger voice during training—but further validation would be needed before judging how dependable it is in clinical practice.