The research, led by Emelie Yonally Phillips and colleagues, examined the implementation of computer-aided detection software, commonly known as CAD, integrated with digital chest X-ray during a community-based tuberculosis active case finding campaign run by Médecins Sans Frontières. Rather than measuring diagnostic accuracy or yield, the study asked a different and often neglected question: how do the people who operate, oversee, and experience these technologies actually perceive them? Between January and April 2023, the team conducted in-depth interviews, focus group discussions, and direct observations with MSF staff, physicians at local health centers, national tuberculosis program personnel, community leaders, and residents who participated in the screening campaign itself.

The technological premise behind the intervention is straightforward but powerful. Digital chest X-rays can be processed by CAD algorithms that flag images suggestive of tuberculosis, prioritizing which patients need confirmatory testing and reducing the dependence on scarce radiologist expertise. In high-burden, resource-constrained settings, this matters enormously. Active case finding campaigns aim to reach people who would never walk into a clinic on their own, and the sheer volume of X-rays generated in such campaigns can overwhelm human reading capacity. CAD promises to widen the funnel, allowing thousands of residents to be screened rapidly while reserving scarce clinical attention for those whose images suggest disease.

Through thematic analysis of the interview and observation data, the researchers identified four major themes that together map the terrain of CAD implementation. The first concerned operational functionality and efficiency, and here the picture was largely positive. Staff reported that the software increased screening capacity, supported clinical decision-making, and streamlined workloads that had previously been bottlenecks. Participants in the active case finding campaign described the process as convenient, an important finding because community screening only works if people are willing to move through it. Speed and simplicity, in other words, are not cosmetic features but determinants of whether a campaign reaches enough people to make a dent in transmission.

The second theme, human resources and role adaptation, revealed a more complicated reality. Staff perceived the redistribution of roles that accompanied CAD adoption as initially challenging. When an algorithm takes over part of the image-reading process, the work of nurses, radiographers, physicians, and data handlers shifts in ways that can unsettle established routines and professional identities. The study’s authors highlight the need for adequate capacity building and technical support during this transition, suggesting that the success of digital health tools depends as much on workforce preparation as on algorithmic performance. Training, in this account, is not a one-off orientation but an ongoing process of helping staff understand what the software does, what it does not do, and how their own judgment fits around it.

Technical feasibility and system design formed the third theme, and it is here that the study offers some of its most practical lessons. Project managers and staff emphasized that understanding CAD’s capabilities and limitations was essential for planning infrastructure, setting thresholds, and compensating for the technology’s inherent weaknesses. CAD systems operate on probability scores, and the threshold at which an image is flagged as suspicious involves a trade-off: lower thresholds catch more true cases but generate more false positives and more confirmatory testing; higher thresholds reduce workload but risk missing disease. Getting this calibration right requires local knowledge of prevalence, health system capacity, and the consequences of error, none of which can be read off a software manual.

The fourth theme addressed community acceptability and ethical considerations, and it complicates any assumption that high-tech solutions are automatically welcomed. Community participants expressed mixed views on the reliability of novel tools, with some skeptical of what a machine could tell them about their health. The researchers stress the importance of aligning the technology with local healthcare needs, a point that carries real weight in a setting like Tondo, where residents may have long histories of interacting with health services that have not always served them well. Trust, the study suggests, is built when communities see that the technology is embedded in care they can actually access, rather than parachuted in as a demonstration of innovation.

Taken together, the findings demonstrate that implementing CAD in tuberculosis active case finding is both feasible and acceptable, but only when supported by careful planning and adaptation to local contexts. The phrase adaptation to local contexts risks sounding like a platitude, but the study gives it concrete content: infrastructure must be planned around the software’s requirements, thresholds must be tuned to local epidemiology, staff roles must be renegotiated with training and support, and communities must be engaged as participants rather than passive subjects of screening. Where any of these elements is neglected, the technology’s promise erodes.

The implications extend well beyond Manila. The World Health Organization has endorsed computer-aided detection as a triage tool for tuberculosis, and donors and national programs are increasingly investing in digital X-ray and AI reading systems across high-burden countries. Yet the gap between a technology that works in principle and one that works in practice is precisely where many implementation efforts fail. The Tondo experience offers a template for closing that gap, emphasizing that sustainable implementation hinges on continuous technical support and on the meaningful involvement of healthcare workers and communities, whose insights help ensure that digital health innovation is ethical as well as effective.

For a disease that kills more people than any other single infection, the stakes of getting this right could hardly be higher. The study’s authors argue that the experience in Tondo offers transferable lessons for deploying CAD in other high-burden, resource-constrained settings, from informal settlements in megacities to remote districts with minimal radiology capacity. What emerges is a portrait of artificial intelligence in global health that is neither utopian nor dismissive: a tool that genuinely expanded screening capacity and eased workloads, but one whose success was earned through negotiation with human systems, professional habits, technical constraints, and community trust. In the fight against tuberculosis, the algorithm is only as effective as the health system and the community that surround it.

Subject of Research: Feasibility and acceptability of computer-aided detection with digital chest X-ray for tuberculosis active case finding in Tondo, Manila

Article Title: Implementing computer‑aided detection for TB active case finding: A qualitative study of feasibility and acceptability in Tondo, Manila

Article References: Yonally Phillips, E., Galvan, M. D. K. P., Min, J., Hewison, C., Duyala, C., Palmado, J. C., Peral, S. N., Camelique, O., Roxas, M. R. C., Duka, M. D., Pardilla, G. F., Recidoro, M. J. C., Castro, R. H., Hossain, F., Huerga, H., & Carnimeo, V. (2026). Implementing computer‑aided detection for TB active case finding: A qualitative study of feasibility and acceptability in Tondo, Manila. PLOS Global Public Health, 6(10), e0007429. https://doi.org/10.1371/journal.pgph.0007429

Image Credits: AI Generated

DOI: 10.1371/journal.pgph.0007429

Keywords: tuberculosis, computer-aided detection, digital chest X-ray, active case finding, artificial intelligence, Manila, Médecins Sans Frontières, global health, qualitative research, screening, digital health, Philippines

The architecture is deliberately simple in concept. The system consists of two independently trained components: an image-level classifier that decides whether a radiograph contains a suspicious lesion at all, and a lesion detector that localizes and marks candidate tumors on the image. The classifier is built on an EfficientNetV2-S backbone, first pretrained on MURA, a large Stanford dataset of musculoskeletal radiographs, and then fine-tuned on BTXRD, a bone tumor X-ray repository, using a two-phase unfreezing schedule that gradually releases layers of the network for retraining. The detector is a YOLOv8 model initialized from weights learned on the COCO object-detection benchmark. Crucially, the classifier can act as a gate: when it judges an image normal, the detector need not run, a design that mirrors the way a human reader first forms a global impression before hunting for specific findings.

The evaluation framework is where the paper makes its most distinctive contribution. The authors argue that summary metrics such as the area under the receiver operating characteristic curve, or AUC, are fundamentally ill-suited to detection tasks because they say nothing about how many false marks a system places on each examination. A detector could post an impressive AUC while littering normal films with spurious annotations that would erode clinician trust in practice. Instead, the team used free-response receiver operating characteristic, or FROC, analysis, the established endpoint for computer-aided detection, and evaluated performance at explicit false-mark budgets that were fixed in advance on validation data before the held-out test set was ever touched.

The data behind the study come from the BTXRD collection, comprising 3,746 radiographs drawn from three sources, of which 1,867 show lesions and 1,879 are normal. These were partitioned into training, validation and test sets of 2,265, 738 and 743 images respectively, with the test set held out completely until all thresholds had been locked. This discipline matters. Many machine learning studies in medical imaging tune their operating points on the very data they then report, inflating performance. Here, every decision threshold was frozen on validation data, and the test results represent a genuine prediction on unseen material.

The headline results are solid if unspectacular, and the authors are refreshingly candid about that. On the test set, the classifier distinguished malignant from normal radiographs with an AUC of 0.933 when initialized with MURA pretraining, compared with 0.910 using standard ImageNet initialization. The benefit of musculoskeletal pretraining was consistent in direction across three random seeds, adding between 0.023 and 0.073 to the AUC, which the authors describe as suggestive of a modest advantage, while noting that the confidence interval at the pre-specified seed included zero. In other words, domain-specific pretraining probably helps, but the evidence stops short of certainty, and the paper says so rather than overclaiming.

Detection performance was assessed at an overlap criterion of IoU at or above 0.3, meaning a predicted mark counts as correct only if it overlaps the true lesion sufficiently. At an achieved rate of 0.027 false marks per normal radiograph, corresponding to a validation budget of 0.02, lesion-level sensitivity reached 0.646. Relaxing the budget to an achieved 0.055 false marks per normal film raised sensitivity to 0.698. At the primary operating point, the system produced 0.40 false marks per lesion-positive radiograph. Sensitivity differed by lesion biology: 0.736 for malignant lesions versus 0.631 for benign ones, a gap that makes intuitive sense given that malignant bone tumors often produce more dramatic radiographic changes.

The matched ablation experiments add further texture. By comparing MURA initialization against ImageNet initialization, and two-phase fine-tuning against full fine-tuning under otherwise identical conditions, the study isolates the contribution of each design choice rather than bundling them into an opaque final system. This kind of controlled comparison is still uncommon in applied medical AI, where papers frequently report only a single best configuration. The leave-one-center-out experiments, however, delivered the study’s most sobering finding. When the smallest of the three data sources was held out, sensitivity held at 0.669, but when the largest source was removed and only 808 training images remained, sensitivity collapsed to 0.287.

That collapse is not a footnote; it is arguably the central lesson of the paper. It demonstrates that cross-source transfer in bone tumor detection is fragile and heavily dependent on the volume and diversity of available training data. A model that performs admirably when trained on nearly 2,300 images from multiple sources can lose most of its diagnostic power when one source disappears from the training pool. For any hospital or research group hoping to deploy such a tool, this is exactly the kind of stress test that should precede real-world use, and the authors explicitly state that cross-source transfer and clinical utility at realistic disease prevalence remain to be established.

The biomimetic framing of the pipeline deserves attention as more than marketing language. Radiologists do not examine every pixel of an X-ray with equal scrutiny; they triage, forming an impression of whether anything is abnormal before committing attention to specific regions. By allowing the classifier to gate the detector, the system embodies a similar division of labor, which carries practical benefits beyond accuracy. Gating can reduce computational cost on normal films and, more importantly, can suppress false marks on images where no lesion exists, directly addressing the false-positive burden that FROC analysis is designed to quantify. The two-stage organization also means each component can be improved, retrained or audited independently, which matters for the regulatory pathways that medical AI systems must eventually navigate.

What makes this study resonate beyond its specific task is its methodological posture. The authors used only publicly available, fully anonymized datasets, received no external funding, and openly disclose the use of generative AI tools for language editing and script drafting while taking full responsibility for the underlying experiments. They resist the temptation to inflate a 0.646 sensitivity figure into a claim of clinical readiness, instead framing sensitivity at fixed false-mark budgets as information that curve-level summaries simply do not convey. In a field where detection systems are often promoted on glossy benchmark numbers, a paper that fixes its thresholds before touching the test set, reports confidence intervals throughout, and shows exactly where its model breaks down when a data source vanishes is a quiet but pointed argument for how medical AI should be evaluated. The bone tumor detector itself is a promising prototype; the evaluation culture it models may prove to be the more consequential contribution.

Subject of Research: Machine learning detection of primary bone tumors on radiographs using FROC evaluation under false-positive constraints

Article Title: A biomimetically inspired staged pipeline for primary bone tumor detection on radiographs: held-out FROC evaluation under explicit false-positive constraints

Article References: Kalejahi, B. K., Khan, S., & Rajabi, M. J. (2026). A biomimetically inspired staged pipeline for primary bone tumor detection on radiographs: held-out FROC evaluation under explicit false-positive constraints. BMC Medical Imaging. https://doi.org/10.1186/s12880-026-02915-x

Image Credits: AI Generated

DOI: 10.1186/s12880-026-02915-x

Keywords: bone tumor detection, radiography, FROC analysis, false positives per image, deep learning, EfficientNetV2, YOLOv8, transfer learning, MURA, BTXRD, leave-one-center-out validation, computer-aided detection

The study, led by Joan C. Vilanova of Clínica Girona and the Institute of Diagnostic Imaging at University Hospital Dr. J. Trueta in Spain, together with colleagues from the medical imaging company Quibim, enrolled 113 men who had been identified with a suspicious lesion on prostate MRI. Every patient in the cohort had an index lesion scored as PI-RADS 3 or higher on the five-point Prostate Imaging Reporting and Data System scale, the standard framework radiologists use to grade the likelihood that a lesion harbors clinically significant cancer. Rather than relying on systematic biopsy patterns, all participants underwent 3-Tesla in-bore MRI-targeted robotic biopsy, a technique in which the needle is guided directly to the suspicious lesion while the patient remains inside the scanner. This design provided a lesion-level reference standard: for each lesion, the researchers knew both what the radiologist saw, what the AI system concluded, and what the tissue actually contained.

The AI system under evaluation, QP-Prostate, analyzed the imaging data and assigned each radiologist-identified lesion a risk classification under two different positivity thresholds. The more permissive threshold counted both Moderate and High risk scores as positive, while the stricter High-only threshold required the algorithm’s top risk category before calling a lesion positive. The clinical question was whether the software could distinguish clinically significant prostate cancer, defined as a Gleason score of 7 or above, from indolent disease or benign tissue. This distinction matters enormously in modern urology: clinically significant cancers warrant active treatment, while low-grade tumors often merit only active surveillance, and avoiding unnecessary diagnosis of insignificant disease is a central goal of prostate cancer screening programs.

The headline numbers tell a story of strong sensitivity purchased at the cost of specificity. Using the Moderate plus High threshold, the AI system correctly identified 92 percent of clinically significant cancers, but its specificity was just 32 percent, meaning it falsely flagged more than two-thirds of lesions that ultimately proved harmless. Tightening the criterion to High-only lowered sensitivity to 79 percent while raising specificity only modestly, to 43 percent. For overall prostate cancer, including Gleason 6 tumors, the pattern was similar: sensitivity of 91 percent with specificity of 43 percent under the permissive threshold, and 80 percent sensitivity with 61 percent specificity under the strict one. In a cohort where clinically significant cancer was highly prevalent at 58.4 percent, these figures describe an algorithm that rarely misses a dangerous tumor but frequently cries wolf.

One of the study’s most technically interesting findings concerns anatomy. The prostate is divided into zones, and the vast majority of cancers arise in the peripheral zone, the outer glandular tissue that is well visualized on the diffusion-weighted and T2-weighted sequences that make up biparametric MRI. The transition zone and central zone, sitting deeper within the gland, present a harder imaging problem because benign enlargement of the transition zone, a near-universal feature of aging male prostates, creates nodules and heterogeneity that mimic malignancy. When the researchers stratified performance by zone, the AI system performed consistently better in the peripheral zone than in the transitional and central zones combined, across both positivity thresholds and both cancer definitions. This zonal dependence echoes a well-known challenge in prostate imaging and suggests that current AI models, like human readers, still struggle most where the imaging appearance of benign and malignant tissue overlaps most heavily.

The exploratory analysis of PI-RADS 3 lesions, the equivocal category where radiologists themselves are least certain, delivered the study’s most cautionary data. In these 18 lesions, the AI system’s sensitivity for clinically significant cancer collapsed to just 33 percent, meaning it missed two of the three significant cancers hiding in this subgroup. Specificity improved only slightly, from 60 percent under the permissive threshold to 67 percent under the strict one. Because PI-RADS 3 lesions represent exactly the gray zone where an objective second opinion from software would be most valuable, this weakness is clinically meaningful. The finding likely reflects both the inherent difficulty of these ambiguous imaging appearances and the composition of training datasets, which tend to emphasize obvious, high-grade lesions rather than the subtle equivocal cases that populate real clinical worklists.

Perhaps the most provocative element of the study is its theoretical simulation of biopsy-sparing, a thought experiment asking what would happen if AI classifications were used to decide which patients could safely skip biopsy altogether. Under the Moderate plus High threshold, 20 of the 113 biopsies, or 17.7 percent, would have been withheld, but 5 of the 66 clinically significant cancers, or 7.6 percent, would have been missed. Under the High-only threshold, the algorithm would have spared 34 biopsies, or 30.1 percent of the cohort, but at the price of missing 14 significant cancers, or 21.2 percent. Among patients without clinically significant cancer, the permissive threshold would have avoided 31.9 percent of biopsies and the strict threshold 42.6 percent. The trade-off is stark and threshold-dependent: every increment in avoided procedures comes with a rising toll of undetected aggressive disease, a currency no clinician can afford to spend lightly.

These results arrive amid intense debate about how AI tools for prostate MRI should be validated and deployed. Many commercial systems report impressive performance metrics derived from retrospective datasets curated for model development, but performance is known to vary substantially across cohorts, scanner vendors, imaging protocols, and disease prevalence. This study’s strengths lie in its real-world framing: consecutive-style recruitment of men already deemed suspicious by radiologists, a robust in-bore targeted biopsy reference standard, and transparent reporting of both thresholds and subgroups. Its limitations are equally instructive. It was a single-centre, retrospective analysis with a small sample, particularly in the PI-RADS 3 subgroup, and the biopsy-sparing simulation was theoretical rather than prospective. The authors are explicit that prospective validation is required before AI-informed biopsy avoidance can be considered clinically safe and effective.

The competitive landscape adds another layer of context. AI-based detection tools for prostate MRI have proliferated rapidly, with several systems gaining regulatory clearances in Europe and the United States, and health systems eager to deploy them amid radiologist shortages and rising imaging volumes. Yet the field has repeatedly seen that high sensitivity in one population does not guarantee generalizability. The present study, notably, evaluated lesions already identified by radiologists rather than asking the AI to find lesions de novo, positioning the software as a risk-classification aid, a computer-aided diagnosis tool in the strict sense, rather than a detection tool. That framing is important: the algorithm’s job was to judge how dangerous a known lesion was, not to spot lesions the radiologist missed, and its performance must be interpreted in that narrower role.

For patients and clinicians, the practical message is one of tempered optimism. The technology demonstrated that it can reliably rule in clinically significant cancer when it flags a lesion as high risk, and its near-perfect sensitivity under the permissive threshold makes it attractive as a safety net that few dangerous tumors escape. But its low specificity means it cannot yet be trusted to rule disease out, and its poor performance on equivocal PI-RADS 3 lesions means the hardest cases remain firmly in human hands. The biopsy-sparing simulation quantifies the dilemma precisely: the choice is not between AI and no AI, but between missing roughly 8 percent versus 21 percent of significant cancers in exchange for sparing 18 percent versus 30 percent of men an invasive procedure. Until prospective trials demonstrate that a particular operating point on that curve is acceptable, the study suggests AI should serve as a decision-support layer alongside radiologists and urologists, informing but never replacing the biopsy conversation. The work, published open access and citable under DOI 10.1186/s12880-026-02863-6, marks a step toward honest, clinically grounded evaluation of prostate AI, and it sets a template for the kind of validation the next generation of these tools will need before they change patient care.

Subject of Research: Clinical validation of an AI-based computer-aided detection system for risk classification of radiologist-identified prostate MRI lesions

Article Title: Clinical validation of an AI-based computer-aided detection system for risk classification of radiologist-identified prostate MRI lesions

Article References: Vilanova, J. C., Martínez-Granados, R., Bazaga, D., Bernabé, E. S., Fuster-Matanzo, A., Llinares-Monllor, C., Polo, A. L., Jimenez-Pastor, A., & Alberich-Bayarri, Á. (2026). Clinical validation of an AI-based computer-aided detection system for risk classification of radiologist-identified prostate MRI lesions. BMC Medical Imaging. https://doi.org/10.1186/s12880-026-02863-6

Image Credits: AI Generated

DOI: 10.1186/s12880-026-02863-6

Keywords: prostate cancer, artificial intelligence, MRI, computer-aided detection, PI-RADS, targeted biopsy, clinically significant prostate cancer, radiology, diagnostic accuracy, machine learning, biparametric MRI, biopsy-sparing

Federated learning is a training framework built around a simple but powerful idea: instead of pooling sensitive medical records on a central server, each participating institution keeps its data locally and trains a shared model on its own machines. In each communication round, a central server distributes the current global model to every client, each client performs local training using its own data, and the server then aggregates the updated parameters to produce a new global model. This approach, known as FedAvg, has attracted intense interest in medical imaging because it promises the statistical benefits of large, diverse datasets while respecting privacy regulations and institutional boundaries that often make data sharing impossible.

The research team, led by Yukihiro Nomura of Chiba University and The University of Tokyo Hospital, tested this idea using two distinct computer-aided detection systems. The first was a three-dimensional U-Net model designed to detect cerebral aneurysms in magnetic resonance angiography images, a task where missing a bulging vessel wall can be catastrophic. The second was a single-shot multibox detector for spotting brain metastases in contrast-enhanced T1-weighted MR images, a critical aid for planning stereotactic radiosurgery. For each task, the team assembled four heterogeneous datasets differing in scanner vendor or institution, totaling 423 MRA images for aneurysm detection and 398 contrast-enhanced T1-weighted images for metastasis detection, with voxel-level annotations available throughout.

The central question was deceptively simple: if a hospital network starts with two institutions and then adds a third and a fourth one at a time, does the detection software actually get better? The researchers hypothesized that it would, but the answer turned out to depend heavily on how the model is retrained when a newcomer arrives. They compared three strategies. From-scratch training reinitialized all model parameters and trained anew on the combined data of all participating clients. Full fine-tuning started from the previously trained global model and updated every parameter. Partial fine-tuning also started from the existing model but froze most layers, updating only a carefully chosen subset, such as the decoder layers of the U-Net or the upper and head portions of the single-shot detector.

The results revealed a clear pattern. For both detection tasks, sequential institution addition combined with fine-tuning generally produced higher median performance gains than training from scratch. Performance improvements tended to accumulate as institutions were added one by one, whereas adding two institutions simultaneously, a condition the team called At-once addition, yielded less consistent improvements. This distinction matters because the simultaneous addition of multiple clients introduces a larger distributional shift in one step, forcing the model to absorb more change at once. When the researchers extended the At-once stage to 200 communication rounds, gains improved in some settings but remained inconsistent across tasks and strategies, suggesting the problem is not simply a matter of insufficient training time.

Perhaps the most striking finding concerns parameter efficiency. In the aneurysm task, full fine-tuning and from-scratch training each updated roughly 0.80 million parameters, while the partial fine-tuning strategies updated only about 0.31 million and 0.08 million parameters respectively. In the metastasis task, the corresponding numbers were approximately 7.95 million for full updates versus 2.66 million and 0.53 million for the two partial strategies. Despite touching a fraction of the network, partial fine-tuning achieved performance changes comparable to full fine-tuning across both tasks. For real-world deployments, this efficiency translates directly into lower computational load and reduced communication overhead, both of which are significant practical constraints when hospitals with varying IT resources participate in a federated network.

The study also confronted an uncomfortable trade-off known in machine learning circles as catastrophic forgetting. When a model adapts to newly added institutions, its performance on the original institutions’ data can decline, which would be unacceptable if the software were already deployed and trusted at those sites. The researchers evaluated this by measuring performance changes on test sets from the initial clients separately from test sets spanning all four institutions. For cerebral aneurysm detection, performance decreases on the initial clients’ test sets were less frequent with full and partial fine-tuning than with from-scratch or centralized training. For brain metastasis detection, however, no strategy showed a clearly distinct pattern, underscoring that the forgetting problem remains task-dependent and unresolved.

Statistical caution tempers the enthusiasm. The team measured performance using the competition performance metric, which averages sensitivity across seven false-positive operating points, and quantified changes as differences from the corresponding initial model. Confidence intervals estimated by paired bootstrap resampling with 1,000 iterations included zero in every evaluated setting on the all-clients test sets, and direct pairwise comparisons among the four training strategies found no statistically significant differences after adjustment for multiple testing. The researchers also ran additional analyses with the FedProx aggregation method, designed for heterogeneous non-IID data, and found broadly similar patterns, suggesting the trends were not an artifact of the FedAvg baseline. Still, the honest conclusion is that the observed benefits are promising tendencies rather than proven guarantees.

The authors frame their setting as a domain-incremental learning problem, a concept from continual learning research in which models must accommodate new data distributions while retaining performance on previously learned ones. This lens highlights both the promise and the limits of the work. The study did not address scenarios such as institutions withdrawing from a network or participating intermittently, nor did it incorporate privacy-enhancing technologies like differential privacy or homomorphic encryption, which would likely be required in production deployments. The clients were also defined by acquisition site or scanner vendor rather than random patient-level splits, capturing realistic heterogeneity in devices and protocols but not the full breadth of differences in patient populations, clinical practice, and annotation policies that large international networks would encounter.

Even with these limitations, the study offers a roadmap for a future in which hospital AI systems improve continuously throughout their operational life. Regulatory frameworks have already approved commercially available detection software bearing CE marks and FDA clearances, and guidelines increasingly recommend multi-institutional data for building robust medical AI. What this research adds is evidence that the network can grow organically, hospital by hospital, with fine-tuned updates that are computationally light and privacy-preserving. As federated learning infrastructure matures, the modest but consistent gains documented here suggest that the collective intelligence of a medical network may indeed be greater than the sum of its parts, provided each new member is welcomed with the right training strategy.

Subject of Research: Federated learning strategies for improving computer-aided detection of cerebral aneurysms and brain metastases as institutions sequentially join a collaborative training network

Article Title: Performance changes in automated lesion detection under federated learning with sequential institution addition

Article References: Performance changes in automated lesion detection under federated learning with sequential institution addition. (n.d.). https://doi.org/10.1007/s11548-026-03795-w

Image Credits: AI Generated

DOI: 10.1007/s11548-026-03795-w

Keywords: federated learning, computer-aided detection, cerebral aneurysm, brain metastasis, magnetic resonance imaging, fine-tuning, deep learning, FedAvg, catastrophic forgetting, medical imaging, multi-institutional data, privacy-preserving AI