AI model combining ultrasound and digital breast tomosynthesis improves specificity in breast cancer triage

The distinction between breast-level and participant-level output deserves careful attention when interpreting the reported performance figures. In clinical breast imaging, a single patient may undergo imaging of both breasts, and each breast may contain multiple findings of varying concern. By assigning a probability score to each breast rather than to each lesion or each patient, the model operates at the unit most relevant to triage decisions: the decision to recall, to biopsy, or to reassure is typically made per breast in screening workflows. This granularity also affects how sensitivity and specificity should be read. A specificity of 0.955 in the pathology-confirmed cohort means that among breasts without malignancy, roughly 96 percent were correctly classified as low risk, leaving about one in twenty benign breasts flagged as suspicious. In a screening population where the vast majority of examined breasts are benign, even small improvements in specificity translate into large absolute reductions in unnecessary recalls, additional imaging, and biopsies, because the denominator of benign breasts is so large.

The choice of a fixed 0.50 decision threshold is a methodological detail with practical consequences. Receiver operating characteristic analysis summarizes performance across all possible thresholds, but deployed triage tools must commit to a single operating point. A threshold of 0.50 balances the model’s tendency toward positive and negative classification, yet different clinical settings might favor different thresholds. A screening environment with limited diagnostic capacity might accept slightly lower sensitivity in exchange for fewer false positives, while a high-risk surveillance program might shift the threshold to prioritize sensitivity. The reported confidence intervals around sensitivity, which overlap with those of the comparator models, suggest that the multimodal approach did not sacrifice cancer detection to achieve its specificity gains, but threshold selection in future deployments will remain a tunable clinical decision rather than an intrinsic property of the model.

The complementary nature of ultrasound and tomosynthesis information explains why their combination outperformed either modality alone and outperformed pairings that included digital mammography. Ultrasound excels at characterizing mass lesions in dense fibroglandular tissue, where mammographic projection is most compromised. It provides real-time assessment of a lesion’s shape, margin, echogenicity relative to surrounding tissue, orientation, and its effect on sound transmission, including posterior acoustic shadowing or enhancement. These features carry independent diagnostic weight. Tomosynthesis, by reconstructing the breast from multiple low-dose projection images, reduces the tissue-overlap artifact that limits two-dimensional mammography and improves the visibility of architectural distortion, spiculated margins, and clustered microcalcifications, features that ultrasound alone cannot reliably detect. Because calcifications and mass morphology represent biologically distinct manifestations of breast disease, a model that can weigh evidence from both streams has access to a broader feature space than either modality provides in isolation.

The architectural choices described in the study reflect established trends in medical image analysis. The use of a modified 2.5D ResNet18 with grouped convolutions for tomosynthesis addresses a specific computational challenge: DBT produces a stack of slices through the breast volume, and processing every slice at full resolution would be prohibitively expensive. The 2.5D approach treats adjacent slices as channels of a single input, allowing the network to capture limited through-plane context without the full memory burden of three-dimensional convolution. Grouped convolutions further reduce parameter counts, which is particularly relevant when training data comprise a few thousand breasts rather than the millions of images available in natural image datasets. The Convolutional Block Attention Module directs the network to spatially salient regions, a mechanism that may be especially valuable when suspicious findings occupy a small fraction of the image volume. Feature-level fusion, in which representations from each modality branch are combined before classification rather than merging raw images, allows each branch to extract modality-specific patterns before the classifier integrates them.

The subgroup findings warrant emphasis because they address the populations where improved triage would matter most. Dense breasts present a well-documented challenge for mammography-based screening: elevated parenchymal density both masks lesions and independently elevates cancer risk, and dense tissue is common, affecting a substantial share of screening-age women. Lesions smaller than two centimeters represent the stage at which diagnosis offers the greatest survival benefit, yet small lesions are precisely where imaging features are least developed and interobserver agreement is weakest. Lower-suspicion BI-RADS categories, typically categories 3 and 4a, generate the majority of benign biopsies and short-interval follow-up recommendations. The observation that the US-DBT model maintained favorable performance across these strata suggests that its specificity gains were not concentrated in easy cases but extended to the ambiguous findings that drive unnecessary work-up.

The retrospective, single-center design imposes limits that the authors appropriately acknowledge. Models trained on examinations from one institution inherit that institution’s scanner fleet, acquisition protocols, patient demographics, and local diagnostic thresholds. Ultrasound is particularly vulnerable to this constraint because it is operator-dependent: the image quality and the selected views reflect the sonographer’s technique, and a model trained on one center’s ultrasound practice may not transfer cleanly to another’s. Tomosynthesis systems from different manufacturers differ in reconstruction algorithms, slice spacing, and dose, and these differences can shift the image statistics on which a deep network relies. External validation across institutions, platforms, and populations is therefore not a formality but a necessary test of whether the learned representations capture disease biology or institutional idiosyncrasy.

Pathology confirmation as the reference standard also shapes what the validation cohort can and cannot show. The independent cohort of 500 pathology-confirmed breasts was enriched for lesions that underwent biopsy, which raises the prevalence of malignancy well above that of an unselected screening population. Higher prevalence inflates positive predictive value and negative predictive value relative to what the same discrimination would yield in routine screening, so the reported PPV of 0.958 should not be read as the proportion of model-flagged breasts that would prove malignant in a general screening setting. What the pathology-confirmed design does provide is a clean test of discrimination: with histopathology available for essentially all included breasts, the AUC and operating-point metrics rest on verified truth rather than on follow-up assumptions. Translating these discrimination estimates into real-world triage benefit will require prospective studies in unselected populations where the model’s flags are adjudicated against subsequent clinical outcomes.

The potential workflow implications extend beyond recall reduction. In many screening programs, the bottleneck is not the initial examination but the downstream diagnostic cascade: diagnostic imaging, biopsies, pathology review, and the patient anxiety that accompanies each step. A tool that refines positive findings at the breast level could function as a prioritization layer, directing limited diagnostic capacity toward breasts with the highest model-estimated risk while allowing lower-risk breasts with benign-appearing findings to follow less intensive pathways. Such prioritization could also reduce reading workload associated with tomosynthesis, which the source evidence notes generates large volumetric datasets that increase radiologist reading time and cognitive load. If the model’s breast-level score can reliably identify low-risk breasts, radiologists might allocate their detailed slice-by-slice review more selectively.

The question of clinician trust and integration into real-time workflows remains open. A retrospective model that receives curated, paired examinations differs from a deployed system that must operate on whatever images arrive, at whatever quality, in whatever sequence. Real-time integration requires that the ultrasound and tomosynthesis examinations be available and linkable at the time of triage, which is not always the case in programs where the two modalities are performed on different visits or by different providers. The authors’ framing of the tool as adjunctive, supporting radiologist-led triage rather than replacing routine assessment, positions it within a human-in-the-loop paradigm in which the model refines rather than dictates decisions. Whether radiologists will calibrate their confidence to the model’s scores, and how such calibration varies with experience and with the visibility of the underlying findings, are questions that only prospective deployment can answer.

The funding landscape described in the source context, spanning national research programs, provincial foundations, and institutional medical-engineering integration initiatives, reflects the multidisciplinary infrastructure that multimodal AI development requires. Combining ultrasound physics, tomosynthesis reconstruction, deep learning architecture design, and clinical radiology expertise within a single research program is resource-intensive, and sustained support across basic and applied streams is often what separates proof-of-concept studies from tools that reach clinical validation. The declared absence of competing interests is relevant here, since the path from an adjunctive triage tool to routine deployment will involve decisions about implementation that should rest on validated performance rather than on commercial incentives.

Finally, the study contributes to a broader evidence base on multimodal integration in cancer imaging. The finding that a dual-modality US-DBT pairing outperformed configurations that included digital mammography suggests that redundancy matters: DM and DBT share the same physical acquisition principle, so their information content overlaps substantially, whereas ultrasound and tomosynthesis interrogate the breast through fundamentally different physical mechanisms. This principle, that complementary modalities add more value than redundant ones, may guide future model design not only in breast imaging but in other oncologic applications where multiple imaging streams are available. The reported AUCs, in the range of 0.93 to 0.94 across two validation cohorts, place the model’s discrimination in the territory where adjunctive clinical use is plausible, provided that the specificity gains survive external validation and that the workflow benefits, fewer false positives and more selective work-up, are demonstrated prospectively in the unselected populations where screening triage actually occurs.

Subject of Research: AI model combining ultrasound and digital breast tomosynthesis improves specificity in breast cancer triage

Article Title: AI model combining ultrasound and digital breast tomosynthesis improves specificity in breast cancer triage

Article References: AI model combining ultrasound and digital breast tomosynthesis improves specificity in breast cancer triage. Original publication (Bibliographic details unavailable.)

Image Credits: AI Generated

DOI: Not provided

Keywords: model, combining, ultrasound, digital, breast, tomosynthesis, improves, specificity, cancer, triage, scientific research