Screening mammograms come with diagnosis labels, not tumor outlines, so models that must say where a cancer is learn from image-level labels alone. Under this weak supervision they reliably find the brightest, most distinctive part of a lesion and miss its faint, infiltrating edge. Radiologists fix this by looking at a second projection of the same breast. Every screening exam already contains two views, the top-down cranio-caudal view (CC) and the angled mediolateral oblique view (MLO). A lesion's distance from the nipple is roughly preserved between them, even though everything else about its appearance changes. We propose to use that one geometric fact as a free training signal. It tells each view's localization map how far along the breast the lesion must extend, and it marks where the two views disagree as a band of boundary uncertainty that the model must resolve. All of the cross-view machinery is used only during training. At inference a small single-view network outlines the lesion on one image, with no paired view and no cross-attention.
The opportunity matters because the field's current best weakly supervised mammography segmenter reaches a Dice overlap of only 0.390 on malignant lesions (Weakly-supervised High-resolution Segmentation of Mammography Images for Breast Cancer Diagnosis). It does so while running two ResNet backbones on 3000×2000-pixel images. The work that does exploit two-view geometry uses it only for representation pretraining with heavy transformers (Geometry-Guided Local Alignment for Multi-view Visual Language Pre-training in Mammography). We ask whether geometry can instead act as a localization supervisor, and whether its benefit survives distillation into a model cheap enough for an ordinary reading workstation.
Core Claim
The claim is that cross-view agreement along the nipple-distance axis supplies two things that image-level labels lack: a coverage target and an uncertainty band. Both come automatically from the paired view, with no human polygons or boxes. Together they close a measurable part of the weakly supervised localization gap, and the gain transfers to a lightweight single-view student.
The mechanism is deliberately one-dimensional. We do not attempt dense pixel matching, because breast compression differs between views and rigid correspondence is unreliable. GLAM's authors flag this fragility of rigid geometric assumptions.
For view $v$, let $S^v$ be the network's saliency map. Let $p^v(d)$ be the maximum of $S^v$ over the strip of pixels at distance $d$ from the nipple, measured along the posterior nipple line. The geometric coverage loss is
$$\mathcal{L}_{\text{geo}} = \sum_{d} \Big[\, \mathrm{maxpool}_\tau\big(p^{\text{MLO}}\big)(d) - p^{\text{CC}}(d) \Big]_+ \;+\; \Big[\, \mathrm{maxpool}_\tau\big(p^{\text{CC}}\big)(d) - p^{\text{MLO}}(d) \Big]_+ ,$$
Here $\tau$ is a tolerance window, in millimetres, that absorbs compression differences, and $[\cdot]_+$ keeps only positive parts. In words: if the other view shows lesion evidence at nipple distance $d$, this view must also show some evidence somewhere in that strip.
This term targets the failure mode, which is that each view collapses onto whatever sub-part is most distinctive in its own projection. Because occlusion by dense fibroglandular tissue differs between CC and MLO, the collapsed parts often differ too. Their union along the distance axis is a better estimate of true lesion extent than either alone.
The same profiles define the uncertainty band. Strips where one view is confident and the other is not mark a "gray zone" of ambiguous boundary pixels. A contrastive term then pulls those pixels toward confident foreground or background feature clusters. That contrastive step is borrowed from EAUWSeg, but there the gray zone is drawn by hand.
Relation to Prior Work
Three papers supply the pieces, and none assembles them.
Weakly-supervised High-resolution Segmentation of Mammography Images for Breast Cancer Diagnosis builds a global-to-local pipeline: a coarse saliency map, high-resolution patches, and a fusion module. It reports Dice 0.390 against 0.325 for its predecessor Globally-Aware Multiple Instance Classifier for Breast Cancer Screening and 0.250 for plain class activation maps. Both systems treat each view independently. They motivate the proposal and are the backbone we modify.
Geometry-Guided Local Alignment for Multi-view Visual Language Pre-training in Mammography shows that matching a patch in one view to its corresponding strip in the other improves BI-RADS prediction by about 2.5% over single-view models. However, it uses this correspondence only inside contrastive vision-language pretraining, never as localization supervision. It also reports that the advantage narrows when fully fine-tuned on small out-of-domain data. We reuse its geometric insight but turn it into a loss on saliency maps, and we test directly whether the narrowing also happens for localization.
EAUWSeg: Eliminating annotation uncertainty in weakly-supervised medical image segmentation shows that an explicit uncertainty band lets weak supervision match full supervision on dermoscopy (86.60% versus 86.25% Dice). Its band, however, costs a hand-drawn pair of polygons per image. We supply the band from the second view instead.
The closest multi-view weakly supervised system is Mammo-Clustering. It fuses CC and MLO with soft attention, reaches AUC 0.828 on VinDr-Mammo, and lowers miss detection rate to 0.294 with about 9.8M parameters. It differs from our proposal in three ways. It imposes no projective constraint. It picks patches from ordinary activation maps, and so inherits their collapse onto sub-parts. It outputs coarse regions rather than pixel boundaries.
Two lines of work constrain the design. Bounding Maps for Universal Lesion Detection shows that a training-only auxiliary branch can raise detection sensitivity by 3.85 points with essentially no inference cost. This is the precedent for keeping the geometry branch out of the deployed model. The negative result in RepSegNet warns that removing its multi-scale blocks cost 8.7 F1 points, most visibly on fine boundaries. Compression can therefore destroy exactly the edges we are trying to recover, and our student design and ablations must check for this.
Research Program
The teacher builds on the high-resolution global-local architecture of Liu et al. It is trained on paired CC/MLO exams with only exam- or breast-level malignancy labels.
Nipple location and the pectoral line, needed to define the distance axis, are estimated by a standard intensity-and-contour heuristic. Its error is measured on a small hand-checked subset so that we know how much geometric noise $\tau$ must absorb.
Training combines three terms: the multiple-instance classification loss, $\mathcal{L}_{\text{geo}}$, and the gray-zone contrastive term. A small sparsity penalty on saliency area counters the obvious failure of $\mathcal{L}_{\text{geo}}$, which is to inflate both maps until every strip is covered.
The student is a single-view encoder-decoder built from structurally reparameterized units. These are multi-branch convolutions during training that are folded into single 3×3 convolutions afterwards, following RepSegNet. The student is distilled from the teacher's refined saliency maps, which are soft targets rather than thresholded masks. At inference it sees one image.
We then analyse three questions. The first is when cross-view collapse is complementary, that is, how often the two views highlight different sub-parts. If they collapse identically, the method has nothing to offer, so this statistic is measured before anything else. The second is how localization gain varies with breast density, the regime where single-view occlusion is worst. The third is how gain degrades as $\tau$ shrinks or the nipple estimate is perturbed.
Baselines
The naive reference point is class activation mapping on a single-view classifier, thresholded into a mask. Liu et al. report Dice 0.250 for it on the NYU dataset. We retrain it ourselves on our datasets, because the NYU dataset is not available to us.
The strongest existing weakly supervised segmenter is Liu et al.'s global-local model. Its reported malignant Dice of 0.390 on NYU is the reference number. No reported weakly supervised Dice was found on Lacuna for CBIS-DDSM, INbreast, or VinDr-Mammo. We will therefore rerun it with the same image-level labels, input resolution, augmentation, and hyperparameter search budget, measured in GPU-hours, as our teacher. Beating it means a statistically significant Dice gain on held-out pixel masks in both CBIS-DDSM and INbreast. A gain on only one dataset does not count.
The strongest multi-view competitor is Mammo-Clustering, which reports AUC 0.805 on CBIS-DDSM and 0.828 on VinDr-Mammo. We compare against it on its own metrics, AUC and miss detection rate, under identical splits. We also extract pixel maps from its saliency to compare Dice. Matching its AUC while improving localization would count as beating it on our question.
The key ablation isolates the geometry itself. It is a two-view teacher trained with unconstrained cross-view consistency: profiles compared without the nipple-distance alignment, equivalently with strips randomly permuted. If this ablation performs as well as the aligned version, the gain comes from simply seeing two views, not from geometry.
A second ablation derives the gray zone from a single view, as the band between two thresholds of the model's own saliency. This tests whether the second view is needed to locate uncertainty.
As upper references, we report a fully supervised U-Net trained on CBIS-DDSM masks. We also cite DUALCORENET's fully supervised 93.69% Dice on INbreast (Dual Convolutional Neural Networks for Breast Mass Segmentation and Diagnosis in Mammography). That figure is computed on pre-cropped regions of interest, so it is a ceiling, not a comparable number.
Why not just…?
Why not just annotate a few hundred masks and train a supervised U-Net? This is the most serious threat, because CBIS-DDSM and INbreast already ship with masks. EAUWSeg's timing (55 versus 10 seconds per image) suggests full masks become impractical at screening scale, but that argument does not show weak supervision is more accurate. We will therefore run a label-budget curve: supervised U-Nets trained on 50, 100, 200, and 400 masks, each evaluated on the same INbreast test masks as our student. We report how many masks the geometric method is "worth." If 50 masks beat it, the honest conclusion is that its value lies only in combination with a small mask budget, and we will test that combination as well.
Why not just fuse the two views at inference, like Mammo-Clustering? Inference-time fusion doubles input cost and requires both views to be available and well registered. The soft-attention fusion also does not by itself produce pixel boundaries. Our experiment includes a variant of our own teacher that uses both views at test time. If it substantially outperforms the distilled single-view student, then training-only geometry leaves accuracy on the table, and the compute claim becomes a trade-off rather than a free gain. We will report it as such.
Why not just fix collapse within one view, with entropy penalties or erasing the confident core? Min-entropy objectives (Min-Entropy Latent Model for Weakly Supervised Object Detection) and reverse attention (Parallel Res2Net-based Network with ReverseAttention for Polyp Segmentation) push maps outward without any second view. Their weakness is that nothing says how far to push, so maps either stay collapsed or bleed into dense tissue. We add both to the Liu et al. baseline under the same tuning budget. If either matches our geometric loss, the second view adds nothing, and the complementarity statistic should already have predicted that outcome.
Proposed Experiments and Decisive Checks
Training uses image-level labels from CBIS-DDSM and VinDr-Mammo, both of which contain paired views. Localization is evaluated on the CBIS-DDSM pixel masks, held out from training, and on INbreast as an external test. VinDr-Mammo boxes are used for box-level hit rate. RSNA-Mammo serves as a classification-only check that AUC is not sacrificed. Metrics are Dice and boundary-band Dice: overlap computed only within a few millimetres of the true edge, the place where the method claims to help. We also report AUC, miss detection rate, and student cost in parameters, GMACs at full resolution, and CPU latency, against the two-backbone Liu et al. model.
The first decisive check is the complementarity measurement on the plain baseline. If fewer than a modest fraction of malignant lesions show different highlighted sub-parts in CC and MLO, we stop and report the negative result.
The second is the aligned-versus-permuted geometry ablation. Geometry must beat mere two-view consistency on boundary-band Dice, or the core claim fails.
The third is the density stratification: gains should concentrate in dense breasts, where occlusion makes the second view informative. A uniform gain across densities would point to a regularization effect, not geometry.
The fourth is the distillation gap. The single-view student should retain most of the teacher's Dice improvement at a small fraction of the baseline's compute. The RepSegNet ablation is our warning that this is not guaranteed for faint edges.
References
- Weakly-supervised High-resolution Segmentation of Mammography Images for Breast Cancer Diagnosis
- Geometry-Guided Local Alignment for Multi-view Visual Language Pre-training in Mammography
- Globally-Aware Multiple Instance Classifier for Breast Cancer Screening
- EAUWSeg: Eliminating Annotation Uncertainty in Weakly-Supervised Medical Image Segmentation
- Mammo-Clustering:A Weakly Supervised Multi-view Global-Local Context Clustering Network for Detection and Classification in Mammography
- Bounding Maps for Universal Lesion Detection
- Efficient medical image segmentation using RepSegNet lightweight reparameterized neural network
- Dual Convolutional Neural Networks for Breast Mass Segmentation and Diagnosis in Mammography
- Min-Entropy Latent Model for Weakly Supervised Object Detection
- Parallel Res2Net-based Network with ReverseAttention for Polyp Segmentation