Image-Based Breast Implant Detection for Mammography Dataset Curation and Near-Real-Time Deployment: Comparing Foundation Models and Task-Specific Convolutional Models

Vasisht Ishwar; Hari Trivedi; Young Seok Jeon; Beatrice Brown-Mulry; Frank Li; Rohan Satya Isaac; Mohammadreza Chavoshi; Judy Wawira Gichoya

arXiv · 2610.03817

Image-Based Breast Implant Detection for Mammography Dataset Curation and Near-Real-Time Deployment: Comparing Foundation Models and Task-Specific Convolutional Models

Vasisht Ishwar·Hari Trivedi·Young Seok Jeon·Beatrice Brown-Mulry·Frank Li·Rohan Satya Isaac·Mohammadreza Chavoshi·Judy Wawira Gichoya

View original source ↗Read source document ↗Export metadata ↓

Abstract

Purpose: To evaluate the performance-feasibility tradeoffs of foundation models (FMs) and task-specific convolutional neural networks (CNNs) trained from scratch for breast implant classification in 2D mammography, with emphasis on suitability for near real-time clinical deployment. Methods: We evaluated four models: two FMs (RAD-DINO and MammoCLIP) and two CNNs trained from scratch for implant prediction (ResNet18 and our lightweight ResNetLite). Using the Emory Breast Imaging Dataset, 5,000 unilateral screening mammograms were used for training/validation and 1,000 manually reviewed unilateral images were held out for testing. For the FMs, global image embeddings from the pretrained encoder were classified using a support vector machine (SVM). The CNNs were trained end-to-end on 2D mammograms, with ResNetLite optimized via grid search over depth and width to balance accuracy and efficiency. Performance was evaluated using AUROC, sensitivity, specificity, accuracy, embedding visualization, and inference-latency. Results: All models demonstrated strong performance on held-out test data (n = 1,000). MammoCLIP achieved the highest AUROC (0.999) with the quickest training time of 493 seconds. RAD-DINO achieved the highest sensitivity (0.980; accuracy 0.989) but had the slowest inference and training times. ResNet18 and MammoCLIP achieved comparable accuracy (0.985). ResNetLite showed no statistically significant difference from ResNet18 (AUROC 0.993; accuracy 0.976) despite using only 1.4% of ResNet18's parameters, and had the fastest inference time. Conclusion: FMs and task-specific CNN models reliably detect breast implants on 2D mammography. Model selection is best guided by deployment context: MammoCLIP for GPU-equipped hospital settings requiring scalable integration, and lightweight CNNs such as ResNetLite for resource-constrained or edge deployments.

Explore related subjects

eess.IV cs.CV q-bio.TO

Keep this discovery

Explore connections, maps & timelines

BibTeX RIS

Vasisht Ishwar, Hari Trivedi, Young Seok Jeon, Beatrice Brown-Mulry, Frank Li, Rohan Satya Isaac, Mohammadreza Chavoshi, Judy Wawira Gichoya. 2026-10-01. Image-Based Breast Implant Detection for Mammography Dataset Curation and Near-Real-Time Deployment: Comparing Foundation Models and Task-Specific Convolutional Models. https://arxiv.org/abs/2610.03817

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

A Simple yet Effective Framework for Pre-treatment and Post-treatment Glioma Segmentation Using Multimodal MRI

Accurate glioma segmentation across pre- and post-treatment MRI remains challenging because treatment alters both image appearance and the target regions. In particular, the resection cavity (RC) is defined only for post-treatment cases, complicating joint learning from pooled data. We present RC-aware nnU-Netv2, a simple yet effective glioma segmentation framework that combines lesion- and boundary-aware single-cohort training, an RC-aware joint learning objective, and treatment-status-guided model routing at inference. During pooled training, the joint objective applies both standard and boundary-weighted binary cross-entropy to all four region channels, while employing Dice supervision for channels with non-empty targets and a controlled false-positive penalty for channels with empty targets across the mini-batch. This design explicitly accommodates the absence of RC in pre-treatment cases. Our approach uses only the standard nnU-Netv2 data augmentation and a single specialized model per case, without additional synthetic tumor generation or multi-model ensembling. RC-aware nnU-Netv2 ranked second in the BraTS 2025 Lighthouse Challenge Task 1 (BraTS-GLI 2025), with segmentation performance comparable to that of the first-place method, which relied on on-the-fly GliGAN-based synthetic tumor augmentation and multi-model ensembling. On the official blind test set, our method achieved mean lesion-wise Dice scores of 0.7878, 0.8709, 0.7923, and 0.8715 and mean NSD@1.0 scores of 0.8309, 0.8712, 0.7980, and 0.8336 for enhancing tumor, RC, tumor core, and whole tumor, respectively. In paired post-treatment analysis, the RC-aware joint model improved mean lesion-wise Dice by 0.042 (95% CI: [0.028, 0.057]) and mean NSD@1.0 by 0.043 (95% CI: [0.028, 0.059]) compared with the joint baseline.

eess.IV↗

arXiv2026

Deficit-Driven Stratified Acquisition for Building STRATA-S12: An Open SAR-Optical Corpus for Self-Supervised Learning

Building a co-registered SAR-optical corpus for self-supervised Earth Observation is above all an acquisition problem, yet the main multi-modal datasets are collected opportunistically, with no per-stratum quota control, so the regions and seasons that cloud cover hits hardest end up under-sampled. We present Deficit-Driven Stratified Acquisition (DDSA), an algorithm that casts the construction of a Sentinel-1/Sentinel-2 corpus as a constrained maximum-coverage problem over a land-cover-driven, hemisphere-corrected (class x season x region) stratification of target countries on all six inhabited continents, and solves it online: acquisition is ordered by marginal gain against a per-stratum deficit that the quality gates and a visual review of every accepted pair keep updating during the run. Three results follow from measuring the algorithm on its own product. First, stratum balance is a property of the deficit-driven ordering rather than of run length: replayed over all 2,928,370 real acquisition attempts, failures included, it populates all 192 strata within 0.05% of the attempts, against 4.03% for uniform round-robin effort and 20.5% for a random order. Second, whether a rejected attempt can be recovered is better predicted by how long its cause persists than by the type of defect: cloud rejections are recovered at eta = 18.8% against 41.4% for nodata, consistent with clouds persisting through a season. Third, the repair loop supplies 39.5% of the released pairs at no measurable quality cost, with 89.8% surviving manual review, but it adds yield rather than balance. The quality guarantee has a price: 2,928,370 acquisition attempts produced 740,746 released pairs. We release the corpus, STRATA-S12, with per-pair metadata sidecars aligned with the STAC and PDS4 standards and the FAIR principles, and a pre-registered linear probe confirms it is fit for self-supervised pre-training.

eess.IV↗

arXiv2026

CoRE: Relating Fine-Grained Chest X-ray Findings in Anatomical Context

Chest X-ray interpretation depends on both the anatomical location of findings and their relationships with other abnormalities. However, relations defined over global finding labels do not explicitly distinguish the same finding across anatomical locations. We propose CoRE(Co-occurrence-guided Region Evidence), a framework that incorporates relationships among findings and anatomical context into fine-grained chest X-ray prediction. It groups anatomical region-finding labels by co-occurrence and uses these groups to guide shared representation learning through visual prompt tuning on a frozen vision transformer. The predictions are supplemented by an anatomically supervised region-evidence stream through region-matched logit residuals. Experiments on MIMIC-CXR with Chest ImaGenome annotations show that CoRE improves region-finding classification over standard and relation-aware prompt-based baselines. Quantitative evaluations on MS-CXR and the Chest ImaGenome gold standard show improved finding localization, supported by qualitative comparisons. CoRE also improves finding-level classification on mapped finding labels in external VinDr-CXR and CheXpert datasets without target-domain adaptation. These results demonstrate the value of jointly modeling finding relationships and anatomical context for fine-grained chest X-ray analysis.

eess.IV↗