PANTHER: AIRI student project wins FLARE 2026, finding tumors in 11 seconds

PANTHER, a model developed by MIPT student Vasily Baranov at the AIRI Summer School, won the FLARE 2026 competition for tumor-detection speed in CT scans. Its solution isn't a race toward architectural complexity — it's three simple ideas about data, ensembles, and filtering.

FLARE 2026: the rules and the task

The competition posed a sweeping task under strict constraints. Find a tumor in any organ (liver, kidney, pancreas, lung, lymph node) in a CT scan within 60 seconds, using no more than 4 GB of video memory, 28 GB of RAM, no internet access, and no external pretrained weights.

Training set: 17,575 annotated CT scans from more than 50 sources, including DeepLesion, KiTS, LiTS, MSD, autoPET, TotalSegmentator, LIDC, and others. Labeling is partial — each dataset only annotates tumors for certain organs, leaving the rest unknown. On top of that: 122 GB of unlabeled CT scans and 172 scans from healthy patients to test for false positives.

Rankings were split evenly: half for speed and efficiency (video memory × inference time per scan), half for accuracy via two metrics. Dice (DSC) measures how closely the predicted tumor mask's volume matches a doctor's annotation (1 = perfect match, 0 = no overlap at all). Normalized Surface Dice (NSD) checks boundaries: what share of the tumor's surface matches within 2 mm. Models slower than 60 seconds get a Dice score of 0 and are disqualified.

PANTHER's three ideas: composition, ensemble, filter

Data composition matters more than architecture. The dataset was heavily imbalanced: 5,700 whole-body CTs, 5,700 chest CTs, but only 154 scans of esophageal cancer, 126 of colon cancer, 79 of endometrial cancer (cancer of the uterine lining), and just 52 of adrenal tumors. Train on everything indiscriminately, and the model becomes an expert on the large sources while only vaguely grasping the rare organs.

The authors built Dataset500 instead: capping the large sources and taking the rare ones in full. The result was 8,762 CT scans from 28 datasets across 11 anatomical groups. Outcome: Dice +0.

060 compared with the full 17,575-scan dataset. By comparison, switching to a residual encoder architecture (with three times more parameters) added only +0.002 — an order of magnitude less.

The authors also coarsened the grid to 2.0×1.5×1.5 mm instead of the recommended 0.78 mm, so a single patch could cover twice as much of the body — large tumors stopped "not fitting." The reverse experiment, with isotropic 1.25 mm, cost −0.078 Dice.

An ensemble of five plain U-Net folds. Five models vote using averaged probability: a voxel is included in the mask only if a majority votes yes. Across 172 healthy CT scans, a single model found a "tumor" in all 172 of them (yes, all of them), while the ensemble stayed silent on 52. Bonus: the ensemble added +0.028 Dice on scans with actual tumors.

A Mahalanobis filter based on mask shape. On the remaining 120 healthy scans, the ensemble still flagged something. The authors identified a telltale signature of "junk": small fragments scattered across the scan, none of them dominant, with a lot of surface area relative to volume. Real tumors look different: dense, compact lesions.

For each mask, they calculate five features: total volume, number of fragments (larger than 3×3×3 voxels), volume of the largest fragment, its share of total volume, and the ratio of surface area to volume. These five numbers form a "portrait" of a typical error. The Mahalanobis distance shows how closely a new mask matches that portrait. If the distance is ≤ 7.0, it's junk, and the mask is erased. Result: false positives dropped from 70% to 8%, with only a 0.004 Dice loss on real tumors.

PANTHER runs in 11 seconds on 4 GB of video memory, outpacing its competitors on speed. It's proof that the right preprocessing and simple ensemble techniques can beat more complex architectures.