Chinese study: AI model combining ultrasound and DBT raised specificity in breast triage without sensitivity loss

Chinese study: AI model combining ultrasound and DBT raised specificity in breast triage without sensitivity loss

Researchers at Sun Yat-sen Memorial Hospital in Guangzhou have developed a deep learning model that merges ultrasound and digital breast tomosynthesis into a single risk assessment per breast. In a pathology-confirmed cohort of 500 breasts, the model reported a specificity of 0.955 — but the results rest, for now, on a single press release, and the authors themselves caution against interpreting the model as a standalone screening tool.

Researchers at Sun Yat-sen Memorial Hospital, affiliated with Sun Yat-sen University in Guangzhou, and collaborating institutions have developed and validated a parallel deep learning framework for breast-level risk classification, based on paired ultrasound (US), digital mammography (DM) and digital breast tomosynthesis (DBT). The study was published in Precision Clinical Medicine, Volume 9, Issue 3, 2026 (DOI: 10.1093/pcmedi/pbag023), and was announced via a press release on September 24, 2026.

The concrete finding is that combining two modalities appears to lift precisely the metric that matters most for triage: the ability to avoid flagging cases as suspicious in error. In the pathology-confirmed cohort, the US–DBT model achieved a specificity of 0.955 (95% confidence interval 0.927–0.975) and a positive predictive value of 0.958 (0.931–0.977), according to the press release. Sensitivity was 0.850 (0.807–0.887) and did not differ significantly from the comparison models. The advantage was therefore not that the model found more cancers — it was that it generated fewer false alarms.

The number that matters most

The model was trained on 2,187 breasts and evaluated in two cohorts: an internal validation cohort of 632 breasts and an independent cohort of 500 breasts confirmed by histopathology as the reference standard. The US–DBT model achieved the highest observed area under the ROC curve (AUC) in both: 0.944 (95% CI 0.926–0.963) internally and 0.934 (95% CI 0.913–0.955) in the pathology-confirmed cohort.

The press release does not provide detailed AUC figures for the single-modality comparison models, so it is not possible to say how large the gap was to, for example, a pure ultrasound or pure DBT model. What can be said from the material is that US–DBT had the highest observed AUC in both cohorts, and that its sensitivity did not differ significantly from the main comparison models. The authors describe the benefit as "improved specificity without a significant loss of sensitivity" — a triage-relevant result, not a breakthrough in how much cancer is detected.

How the model is built

The architecture consists of modality-specific branches that process ultrasound and DBT separately before merging them. On the DBT side, a modified 2.5D ResNet18 with grouped convolutions is used — an adaptation to tomosynthesis's stacked image strips, which differ from ordinary two-dimensional mammography images. A CBAM attention module (Convolutional Block Attention Module) is used to weight which image regions influence the assessment, and the features from the two branches are fused at the feature level before a multilayer perceptron (MLP) performs classification.

Each model produces a probability score per breast, with a prespecified threshold of 0.50. This means the output is a binary risk assessment per breast — not a diagnosis per lesion, and not a replacement for the radiologist's reading.

Why specificity carries weight in triage

In a triage setup where the model is meant to help prioritize which breasts should receive follow-up, the cost of false positives is not just unnecessary follow-up examinations but also patient anxiety and strain on capacity. A model that keeps sensitivity stable while specificity rises will in practice flag fewer breasts as "suspicious" without missing cancers — and this is what the authors point to as the model's most consistent advantage.

It is worth noting that this is classification at the breast level, not the patient level or the lesion level, and that the study uses histopathology as the gold standard. This is a methodological strength in the external cohort, but it also means the results apply to populations in which biopsy has already been performed — not a general screening population.

The authors' own caveats

The authors stress that the model is not intended as a standalone screening or diagnostic system, and that prospective, multicenter validation is needed before clinical implementation. They also note that the retrospective, single-center design limits generalizability: broader use will depend on external validation across institutions, imaging platforms and patient populations, as well as prospective testing of real-time workflow and clinicians' trust in the tool.

In other words: the results are promising internally, but there is as yet no documentation of how the model performs in a different department, on a different ultrasound machine, or in a prospective clinical setting.

Sourcing: all figures come from a single press release

AIMag has not had access to the peer-reviewed article itself. All figures and quotations in this story come from the press release distributed via news-medical and Newswise, dated September 24, 2026. The two available copies are textually identical and derive from the same release, so they do not constitute independent confirmation of the findings. The press release was issued via West China Hospital of Sichuan University, but the relationship between that institution and the Sun Yat-sen team is not explained in the sources. The article was authored by Tan Y. with co-authors, according to the reference in the press release.

The reported performance figures should therefore be read as what the study itself reports, not as independently verified results. It is also not known from the available material how large the differences to single-modality baseline models were in practice.

What remains

Before the model can be used clinically, it must be validated prospectively at multiple centers, across different imaging platforms and patient groups. The authors also point to the need to test how such a tool actually functions in a real-time screening workflow — including whether clinicians trust it enough to act on its assessments. Until such studies exist, the reasonable picture is a promising triage tool with a clear mechanism (merging complementary modalities) and a concrete but still unconfirmed promise of better specificity.