Object detectors learn by deciding, for every ground-truth box, which of the model's candidate predictions count as "positives" to be trained toward it. The usual rule scores candidates by Intersection over Union (IoU), the overlap area divided by the combined area. For small boxes this rule is brittle: a shift of a few pixels can drop IoU from high to zero, so small objects get fewer and worse positive examples during training. We propose that in chest radiographs this creates a hidden, population-dependent training sampler. A radiograph is a flat projection of a 3D chest, so a lesion's apparent size and outline depend on the patient's body, through thoracic dimensions, overlying soft tissue, and age-related anatomy. If lesions in one group project smaller or more crowded, an IoU-based matcher trains on them less, and a gap in detection accuracy between groups follows without any demographic shortcut in the features.
The proposal has two parts. First, we measure this "assignment gap" on VinDr-CXR, the open chest X-ray dataset with radiologist bounding boxes, stratified by sex and age. Second, we test whether replacing IoU with geometric similarity matching closes the gap, and whether it does so for the reason we claim. Geometric similarity matching scores candidates by normalized center distance plus width, height, and aspect-ratio agreement, and it has been shown to rescue tiny objects in aerial detection. We isolate the mechanism against cheaper fixes that equalize only the number of positives, or that simply reweight groups.
Core Claim
Part of the subgroup disparity in radiograph lesion detection is produced by the label-assignment step. It is mediated by projected lesion size and crowding, and a matching score that stays continuous across scale reduces this part. It does so without the accuracy loss that explicit fairness penalties cause. The claim is falsifiable in three ways. If per-group differences in positives-per-box are negligible, the premise fails. If disparity persists unchanged after conditioning on box size, the mechanism is wrong. If a size-agnostic count equalizer matches geometric matching, geometry is not the active ingredient.
We summarize assignment per group $g$ with one quantity. Let $n_g$ be the average number of positive training candidates per ground-truth box in group $g$, and $q_g$ the average matching quality (IoU) of those positives. The assignment gap is the pair $(n_{g} - n_{g'},\; q_{g} - q_{g'})$ for two groups $g, g'$, and the audit reports it as a function of box area.
Where This Sits in Prior Work
The method comes from small-object detection. Revisiting DETR for Small Object Detection via Noise-Resilient Query Optimization replaces IoU in proposal matching with a Pairwise-Similarity metric that combines normalized center distance and relative width and height differences. It reports 32.1% AP on VisDrone and shows in ablations that center-distance-only metrics lose 1–3% AP in cluttered scenes. COXNet uses a related GeoShape score for dynamic label assignment on drone person detection. We reuse both scores unchanged; our contribution is to measure what they do to who gets trained on, which neither paper considers.
AMSP-UOD adds aspect-ratio similarity to non-maximum suppression, the step that deletes duplicate boxes, so that adjacent true objects are not suppressed in dense underwater scenes. BEVStereo scales its suppression radius by object size. Both use a single threshold for all inputs. We use them as a secondary probe of whether post-processing suppresses valid findings unevenly across groups.
The population-bias motivation comes from classification audits. Risk of Bias in Chest X-ray Foundation Models shows that frozen features separate patients by sex and race. It finds drops of 6.8–7.8% for female "no finding" and 10.7–11.6% for Black patients' pleural effusion, and it attributes the gap to feature shortcuts. Our hypothesis is complementary. In detection, a gap can arise even with neutral features, through the training sampler.
Fairer AI in ophthalmology supplies the design principle. Grounding predictions in localized pathology raised mAP to 92.0% on OculoScope while narrowing the sex gap. It did this through attention, on a non-projective modality. We test the same principle through explicit box geometry on a projective one.
Improving the Fairness of Chest X-ray Classifiers constrains the design. Nine debiasing methods traded overall accuracy and calibration for parity, while simple balanced resampling matched them on worst-group performance. Balanced resampling is therefore a mandatory baseline. The same paper also warns that label noise can vary by age.
Projection loss enters as the mechanism rather than the thing we fix. FVNet documents that projection collapses small targets into a few pixels. AbdCTBench shows that a projected body representation yields sex-dependent error, but it involves no localization. We do not attempt 3D recovery.
Research Program
We train one detector family, a DETR-style detector matching the NRQO setting, together with a one-stage anchor detector for generality. Training uses the VinDr-CXR split of 15,000 training and 3,000 test images described in VinDr-CXR. Patient sex and age are read from DICOM headers, and we report coverage per split before any analysis. Race is not available in this cohort, so we do not claim anything about it.
Radiologists' boxes are fused by a fixed rule shared across all arms. We also report results under a second fusion rule, because annotator disagreement could itself differ by group.
The audit logs, at every training step, each ground-truth box's positive count and match quality under the active assigner. It then fits $n_g$ and $q_g$ against box area, aspect ratio, and local box density. A mediation analysis asks how much of the per-group AP gap is explained by projected box size and crowding.
The intervention arms swap only the assignment score, holding backbone, schedule, augmentation, and data identical. Separately, we apply NMS-Similar and size-aware suppression at test time and measure per-group rates of true positives deleted by suppression.
Baselines
The naive reference is the same detector with standard IoU assignment and standard non-maximum suppression, trained with plain empirical risk. Every gap and gain is measured against it.
The strongest published number on this benchmark in our file is CXR-RefineDet, which reports mAP 0.1686 on VinDr-CXR at 512×512. We reproduce its resolution and evaluation protocol, so that our IoU baseline is anchored to a published figure rather than to a weak internal run. No reported per-sex or per-age detection result found on Lacuna for VinDr-CXR, so the subgroup baselines are our own reproductions.
Balanced ERM upsamples images so that the sex and age bins are equal in each epoch. Improving the Fairness of Chest X-ray Classifiers found it matched or beat GroupDRO, JTT, and adversarial methods on worst-group performance, which makes it the strongest competing fix. It gets the same epochs and a matched tuning budget of equal trial count over the sampling ratio. We also run it combined with geometric matching, since the two could be additive.
One group-robust optimizer, GroupDRO, is included to confirm the accuracy-parity trade-off in detection. It is given the same tuning budget.
Three ablations isolate the key ingredient. A count-equalized IoU assigner, which takes the top-k candidates per ground-truth box regardless of IoU, fixes $n_g$ but not match quality or shape. A center-distance-only score, in the style of the DotD and NWD comparisons in NRQO, removes shape. The Pairwise-Similarity and GeoShape scores are then used as published. The balance coefficient between the location and shape terms is swept over the same grid for all arms, because NRQO reports sensitivity to its $\alpha$.
Beating the field means two things. Geometric matching must reduce the worst-group AP gap relative to the IoU baseline with bootstrap confidence intervals excluding zero. It must also do so without lowering overall mAP below the IoU baseline, and do better than both Balanced ERM and count equalization on at least one of those two axes.
Why not just…?
Why not just collect more diverse data? More data does not change how the matcher treats a small box. Every additional small lesion is still under-assigned under IoU. The foundation-model audit in Risk of Bias in Chest X-ray Foundation Models found larger demographic separation in a model trained on broader data than in a simpler baseline. Our data-fraction sweep, which trains on 25%, 50%, and 100% of the training set, directly tests whether the assignment gap shrinks with scale.
Why not just rebalance groups (Balanced ERM)? This is a real threat. If it closes the AP gap as well as geometric matching does, our claim that assignment is the cause weakens. Rebalancing changes how often a group's images are seen, but not how many positives each of its small boxes receives per view. The audit measures exactly this: we report $n_g$ and $q_g$ under Balanced ERM. If rebalancing equalizes AP while leaving the per-box gap intact, the two act through different routes and should combine.
Why not just add a fairness penalty or use GroupDRO? In chest X-ray classification these methods bought parity by lowering accuracy and worsening calibration for every group (Improving the Fairness of Chest X-ray Classifiers). Geometric matching instead targets a source of error that hurts accuracy everywhere. We include GroupDRO to check whether the trade-off repeats in detection.
Why not just give every box the same number of positives? This is the sharpest alternative, because it fixes the counting half of the mechanism for free. We run it as the count-equalized ablation. Our prediction is that it helps for isolated small boxes but not in crowded regions, where extra positives get matched to a neighboring finding. Geometric shape terms were designed for that crowded case, per the NRQO ablations. If count equalization matches geometric matching across density strata, we will report that geometry is unnecessary.
Why not just use the 3D data? Chest radiography is a planar screening modality with no paired volume in VinDr-CXR. The question is how to train well on projections as they are deployed.
Decisive Experiments
The first check is the audit itself. Under IoU assignment, do $n_g$ and the fraction of boxes with zero positives differ between sexes and age bands, and does that difference vanish when boxes are matched on area? If groups do not differ in projected box size, the mechanism is absent and we report a negative result.
The second check is a controlled size intervention. We rescale images from one group so that its box-size distribution matches the other's, retrain, and ask whether the AP gap moves in the predicted direction. This separates a projected-size effect from an appearance effect.
The third check is the arm comparison. For each assigner we report overall mAP, per-group AP, worst-group AP, and gap, with calibration of box confidence per group. Results are stratified by box area and local box density, so that the count-equalized, center-distance, and full-geometry arms can be told apart in exactly the regime where they should differ.
The fourth check tests suppression. Test-time NMS-Similar is applied on top of each trained model to see whether any remaining gap comes from suppression rather than training.
The study succeeds as science whichever arm wins. It will say whether detection disparities in projected medical images start at the matcher, and which property of the matching score removes them.
References
- Revisiting DETR for Small Object Detection via Noise-Resilient Query Optimization
- COXNet: Cross-Layer Fusion With Adaptive Alignment and Scale Integration for RGBT Tiny Object Detection
- AMSP-UOD: When Vortex Convolution and Stochastic Perturbation Meet Underwater Object Detection
- BEVStereo: Enhancing Depth Estimation in Multi-View 3D Object Detection with Temporal Stereo
- Risk of Bias in Chest X-ray Foundation Models
- Fairer AI in ophthalmology via implicit fairness learning for mitigating sexism and ageism
- Improving the Fairness of Chest X-ray Classifiers
- FVNet: 3D Front-View Proposal Generation for Real-Time Object Detection from Point Clouds
- AbdCTBench: Learning Clinical Biomarker Representations from Abdominal Surface Geometry
- VinDr-CXR: An open dataset of chest X-rays with radiologist's annotations
- CXR-RefineDet: Single-Shot Refinement Neural Network for Chest X-Ray Radiograph Based on Multiple Lesions Detection