A team of researchers from Qatar University and the University of Hyogo has now taken a direct swing at that problem. In a study published in Multimedia Tools and Applications, they introduce TomoMamba, a two-stage deep learning framework that, according to the authors, is the first application of selective state-space models to inter-slice feature propagation in DBT. The work arrives at a moment when the Mamba architecture, a relative newcomer to the deep learning scene, has been making waves for its ability to model long sequences with linear computational cost, in contrast to the quadratic cost of transformer attention.

The framework’s design mirrors the clinical workflow of breast screening. Stage 1 acts as a high-throughput triage system: a dual-branch 2.5D convolutional network processes the craniocaudal and mediolateral oblique views of each breast through a shared ResNet-18 backbone, aggregates information across slices with a learnable attention mechanism, and fuses the two view-level representations into a single verdict of normal or suspicious. On the validation partition of the public BCS-DBT dataset, this screening stage achieved a breast-level area under the receiver operating characteristic curve of 0.97, correctly filtering out 94.5 percent of normal breasts while identifying 36 of 41 suspicious ones.

Only breasts flagged as suspicious proceed to Stage 2, where the real architectural novelty lives. A ResNet-18 backbone with a lightweight feature pyramid network extracts features from each of the fifteen representative slices retained per volume. Then, at every spatial position in the feature map, the model extracts a fifteen-element sequence of slice-wise features and feeds it through a Mamba selective state-space model. The crucial distinction from earlier recurrent approaches is that Mamba’s state-transition matrices are input-dependent: the amount each slice contributes to the running hidden state varies with the content of that slice. A lesion that activates strongly on two or three adjacent slices therefore leaves a depth summary that carries information about where along the depth axis the signal appeared, something slice-independent methods discard entirely and full 3D convolutions capture only at substantial memory cost.

The Mamba output is folded back into the per-slice features through a gated residual connection, a sigmoid gate that makes per-channel keep-or-suppress decisions conditioned on the depth context. Introspection across 72 validation volumes showed the gate is channel-selective rather than slice-selective: at lesion positions a mean of 40 of 128 channels fell outside the neutral band, against 23.5 at background positions, a contrast of 1.70 times with a large effect size. Downstream, an anchor-free CenterNet-style detector predicts heatmaps, box sizes and offsets directly from the Mamba-enhanced features, and classification is performed on small feature patches cropped at the detected lesion centers rather than pooled globally. That cropping choice proved essential: a pooling-only baseline oscillated between all-cancer and all-benign predictions without ever converging on the tiny training set of 200 biopsied volumes from just 101 patients.

The evaluation rests on BCS-DBT, hosted on The Cancer Imaging Archive and collected at Duke University Hospital, which remains the only public DBT dataset combining biopsy-confirmed pathology labels with lesion-level bounding-box annotations. The class distribution is extreme, with normal cases outnumbering cancers by more than 100 to 1, which is precisely why the authors split the problem in two. On the validation partition, Stage 2 attained a patient-level AUC of 0.80 with a lesion detection rate of 72.0 percent. On the untouched test partition, however, classification fell to 0.62, an interval that spans chance, while detection degraded only modestly to 62.5 percent. The authors are unusually candid about this: they present lesion localization, not benign-versus-cancer discrimination, as Stage 2’s demonstrated contribution.

Chained together, the two stages reached a breast-level AUC of 0.93 on validation and 0.92 on the held-out test set, forwarding just 14 percent of breasts for expensive diagnostic analysis while losing two of 20 cancers at screening. The authors warn that this cascade figure reflects the screening stage’s normal-versus-suspicious separation rather than genuine cancer discrimination, and that roughly one cancer in ten would be missed if the cascade were used as a standalone filter. They position the system as a reader-support tool to be used alongside, not in place of, radiologist interpretation.

The controlled baseline comparisons are where the architectural argument gets its sharpest test. Under identical data, preprocessing and hyperparameters, TomoMamba was pitted against a bidirectional GRU, a transformer variant and a slice-independent baseline, all matched to within one percent on parameter count. On the validation partition, the cross-slice architectures were not separable by classification AUC, and the authors decline to rank them on that metric. But at matched false-positive rates, TomoMamba localized more lesions than every alternative on validation, and on the test partition it achieved the highest patient-level AUC and the lowest Brier score of the five configurations. Most tellingly, removing the gated cross-slice residual dropped patient-level AUC from 0.798 to 0.548, the only architectural ablation that reached statistical significance at 0.009.

The team also confronted a subtle methodological hazard head-on. Because radiologist annotations guided which slices were retained during preprocessing, the model could in principle have exploited the annotated slice’s position as a hidden marker. The authors tested this directly: detection rates showed no significant difference depending on where the annotated slice sat in the sequence, and under a fully annotation-independent selection procedure, patient-level classification changed little, falling from 0.798 to 0.748 on validation. Since annotations do not exist at deployment, they treat the annotation-independent figures as the deployment-relevant results. Failure analysis added further nuance: the dominant failure mode was classification rather than localization, with most misclassifications clustered within 0.03 of the decision threshold, suggesting a calibration problem rather than a detection one.

The limitations are stated with rare honesty. The study is single-institution, single-scanner, and underpowered for benign-versus-cancer classification with only 101 biopsied training patients; no published DBT method has yet demonstrated cross-site transfer. The much-hyped efficiency of state-space models also did not materialize at this scale: at fifteen slices per volume, Mamba’s sequential scan was actually slower than both the recurrent and attention baselines, though its linear scaling should win out on deeper acquisitions. Still, the significance of the work lies in its disciplined template for evaluating new architectures on small clinical datasets, reporting all three official partitions, quantifying seed variability, and refusing to claim more than the data can support. As larger annotated DBT cohorts emerge, the question of whether selective state-space propagation becomes the backbone of volumetric cancer screening will finally get the data it deserves.

Subject of Research: Deep learning with selective state-space models for cross-slice analysis in digital breast tomosynthesis for breast cancer screening and diagnosis

Article Title: TomoMamba: a two-stage model with state-space cross-slice propagation for breast cancer diagnosis in digital breast tomosynthesis

Article References: Soliman, S., Zafari, Y., Rashed, E. A., & Mabrok, M. (2026). TomoMamba: a two-stage model with state-space cross-slice propagation for breast cancer diagnosis in digital breast tomosynthesis. Multimedia Tools and Applications, 85(10), Article 776. https://doi.org/10.1007/s11042-026-21943-x

Image Credits: AI Generated

DOI: 10.1007/s11042-026-21943-x

Keywords: digital breast tomosynthesis, breast cancer, Mamba, state-space models, deep learning, lesion detection, computer-aided diagnosis, BCS-DBT dataset, medical imaging AI, CenterNet, screening cascade, feature propagation

The findings arrive at a critical moment. In 2020, the World Health Organization launched its global strategy to eliminate cervical cancer as a public health problem, built on the so-called 90-70-90 targets: 90 percent of girls fully vaccinated against human papillomavirus by age 15, 70 percent of women screened with a high-performance test in their lifetime, and 90 percent of women with cervical disease receiving appropriate treatment. While considerable attention has focused on vaccination and screening coverage, the third target—treatment—has received far less scrutiny. Yet a screening program that identifies precancerous lesions but fails to shepherd women through diagnosis and treatment saves no one. Public health researchers call this the screening cascade, and the losses at each step, from a positive screen to colposcopy attendance to diagnostic confirmation to definitive treatment, are often dramatic, particularly in low- and middle-income countries where the majority of cervical cancer deaths occur.

Leading the new analysis, Xinhua Jia of Chongqing University Cancer Hospital and the National Cancer Center in Beijing, together with colleagues including senior author Youlin Qiao, set out to answer two questions that have frustrated program designers for years. First, how effective are interventions at improving follow-up completion among screen-positive women? Second, and more subtly, does the effectiveness of those interventions depend on what kind of intervention is delivered, or on where in the cascade the intervention is applied? The distinction matters enormously for resource allocation: if SMS reminders work as well as intensive patient navigation, cash-strapped health systems in low- and middle-income settings could favor cheaper digital tools. But if effect sizes shrink predictably at later cascade steps, a single-point intervention may never be sufficient.

The team searched PubMed, Embase, CINAHL, and the Cochrane Library from their inception through 31 March 2026, registered the review prospectively with PROSPERO under identifier CRD420251031531, and followed PRISMA 2020 reporting standards. From 33 studies that met inclusion criteria, 25 studies covering 142,298 women in 11 countries across six continents provided data for the primary meta-analysis. The investigators used Hartung–Knapp–Sidik–Jonkman random-effects models as their primary statistical approach, a method favored for its better control of false-positive findings when the number of studies is modest and heterogeneity is high. Subgroup analyses examined study design, intervention type, country income level, the presence of a digital component, and, in exploratory fashion, the position of each intervention along the screening cascade. Publication bias was assessed with Egger’s regression test, trim-and-fill analysis, and the PET-PEESE method, and certainty of the evidence was graded using the GRADE framework.

The headline result was striking: structured interventions were associated with significantly higher odds of follow-up completion, with an odds ratio of 2.03 (95 percent confidence interval 1.71 to 2.41). In practical terms, a woman in an intervention arm had roughly twice the odds of completing her follow-up as a woman receiving usual care. Heterogeneity was substantial, with an I-squared value of 83.1 percent, indicating that the true effect likely varies considerably across settings and populations—a reality captured by the wide prediction interval of 1.11 to 3.73, which suggests that in some settings the effect could be nearly negligible while in others it could be nearly fourfold. Importantly, the direction of effect was consistent across study designs, with subgroup odds ratios ranging from 1.80 to 2.21 whether the underlying evidence came from randomized controlled trials or observational studies.

The authors did not shy away from the field’s uncomfortable methodological realities. Egger’s test indicated substantial publication bias, with a P value below 0.001, meaning small studies with null results are likely underrepresented in the published literature. When the team applied trim-and-fill and PET-PEESE adjustments, the pooled estimate attenuated to an odds ratio between 1.73 and 1.88, and a restricted analysis of the 15 studies judged to be at low or moderate risk of bias yielded an odds ratio of 1.87 (95 percent confidence interval 1.55 to 2.26). Even after these corrections, the signal remained robust and clinically meaningful. The authors translated the adjusted estimate into absolute terms: for a program with a baseline follow-up completion rate of 30 percent—a sobering but realistic figure in many high-burden settings—an intervention with an odds ratio of about 1.8 would lift completion to roughly 44 percent, an absolute gain of about 14 percentage points.

Perhaps the most consequential finding concerned what did not predict effect size. Subgroup differences by intervention type, by the presence of a digital component, and by country income level were all statistically nonsignificant, with P values exceeding 0.25. In other words, the analysis found no clear evidence that expensive, labor-intensive navigation programs outperform lightweight SMS reminders, nor that interventions work better in wealthy countries than in resource-constrained ones. For ministries of health weighing competing investments, this is an encouraging message: what appears to matter is not the sophistication of the delivery modality but the presence of some structured, deliberate mechanism connecting a screen-positive woman to her next appointment.

Where the effect did vary was along the cascade itself. In exploratory analyses, effect sizes differed significantly by cascade-step position, with a P value of 0.006. Interventions were most effective at getting women in the door for colposcopy, where the odds ratio reached 2.35, and least effective at achieving full diagnostic resolution, where the odds ratio fell to 1.46. The pattern suggests a gradient in which structured support is powerful at overcoming the logistical and informational barriers of a single next step—scheduling a colposcopy, arranging transport, reminding a patient of a date—but loses traction as the pathway lengthens and multiplies. Reaching diagnostic resolution may require several consecutive hurdles: biopsy, histopathological confirmation, and a treatment decision, each of which presents its own opportunity for attrition. The implication is sobering for program designers who treat follow-up as a single problem to be solved once.

The GRADE assessment rated the overall certainty of the evidence as very low, a reminder that much of the underlying literature rests on observational designs with inherent vulnerabilities to confounding and selection effects. The authors are candid about this limitation and about the substantial between-study heterogeneity, and their conclusion is correspondingly measured: the association between structured interventions and improved follow-up is consistent, but the field lacks the rigorous, randomized evidence—particularly from low- and middle-income settings—needed to prescribe precisely which intervention packages should be deployed at which cascade steps. They call explicitly for layered, step-specific interventions that prioritize the distal, most leak-prone portions of the diagnostic pathway, where gains appear hardest to achieve and where the 90 percent treatment target will ultimately be won or lost.

With the 2030 deadline for the WHO elimination targets fast approaching, the study reframes the treatment gap not as a monolithic shortfall but as a series of discrete, addressable losses. Its quantitative core—an adjusted odds ratio near 1.8 and an absolute gain of roughly 14 percentage points at low baseline completion rates—gives policymakers a realistic benchmark for what structured follow-up support can deliver, while its cascade-step analysis warns against complacency at the final mile. As cervical cancer continues to claim hundreds of thousands of lives annually, most of them in countries with the weakest screening infrastructure, the message is clear: finding the lesions is no longer the hardest part. Getting women through the door, and then through every door after that, is where elimination will be decided.

Subject of Research: Effectiveness of structured interventions in improving follow-up completion along the cervical cancer screening cascade toward WHO 90-70-90 elimination targets

Article Title: Closing the cervical screening cascade gap toward WHO 90-70-90 elimination targets: A systematic review and meta-analysis of 25 studies in 11 countries

Article References: Jia, X., Da, X., Sun, R., Shi, J., Gao, C., & Qiao, Y. (2026). Closing the cervical screening cascade gap toward WHO 90-70-90 elimination targets: A systematic review and meta-analysis of 25 studies in 11 countries. BMC Medicine. https://doi.org/10.1186/s12916-026-05221-z

Image Credits: AI Generated

DOI: 10.1186/s12916-026-05221-z

Keywords: cervical cancer, WHO elimination targets, screening cascade, meta-analysis, patient navigation, colposcopy, follow-up completion, HPV screening, low- and middle-income countries, publication bias, BMC Medicine, global health