AI-assisted tracking sharpened longitudinal brain metastasis follow-up
Tracking brain metastases over time is often harder than measuring them at a single visit. Once stereotactic radiotherapy begins, radiologists must compare current images against a stack of prior studies, then decide whether a lesion has shrunk, stayed stable, or quietly begun to grow again. In this study, a PACS-integrated artificial intelligence prototype lesion tracking tool was built to make that process more systematic, and the signal it produced was clinically meaningful: the workflow improved follow-up of lesions and identified progression that manual review missed.
The investigation focused on patients with brain metastases who underwent stereotactic radiotherapy. Manual review relied on two board-certified neuroradiologists who displayed the current scan alongside up to seven prior studies and measured orthogonal lesion diameters by hand. The AI-assisted workflow used a custom hanging protocol to automatically select, display, and three-dimensionally register contrast-enhanced T1 MRI sequences across as many as eight studies. A Scalable and Transferable U-Net, trained on the BraTS METS 2023 brain metastasis dataset, generated segmentations from which lesion diameters were extrapolated. Radiologists then revised those measurements when needed, creating a hybrid human-machine workflow rather than a fully autonomous one.
Missed lesions still reached progressive disease
The cohort included 40 patients and 158 MRIs. Among 353 true lesions, 59 were detected by only one neuroradiologist and 15 by only the second reader. The disagreement was not merely academic. Of the 74 lesions that differed between readers, 14 went on to complete response, 4 to partial response, 39 remained stable disease, and 12 developed progressive disease. In practical terms, a subset of lesions that were difficult to identify consistently at the reading stage later proved to be biologically active disease.
A one-sentence truth sits inside those numbers: lesions missed on one read were not uniformly benign. Some eventually declared themselves as progressive disease.
The median Dice similarity coefficient for the STU-Net segmentations was 0.771, a result that indicates substantial overlap but also leaves room for error in a setting where small geometric changes matter. In brain metastasis surveillance, a few millimeters can separate stability from treatment failure, especially when clinicians are trying to decide whether local therapy has held or whether a lesion requires renewed intervention. The study does not claim the AI model eliminated error; it shows something more useful and more realistic for daily care: the system improved longitudinal assessment enough to surface disease that conventional side-by-side review could overlook.
Reader variability is the central clinical problem
The strongest signal in the study may be the degree of reader disagreement. Manual diameter measurements showed high agreement overall, with intra-class and Spearman correlation coefficients of 0.959 and 0.961, respectively. Those values suggest that, on average, the readers were highly concordant. Yet the lesion-level data tell a different story. High aggregate correlation can coexist with clinically relevant misses in individual lesions, especially in patients with multiple metastases spread across several scans and multiple treatment time points.
That is exactly where automated longitudinal registration can change the reading task. When the software aligns serial MRI studies and presents them in a consistent framework, the neuroradiologist is no longer reconstructing the disease course from memory and manual scrolling alone. The workflow becomes less dependent on the reader’s ability to mentally register lesion position and contour across time. In a disease defined by multiplicity and motion over time, that kind of assistance is not a convenience feature; it directly addresses the main source of interpretive friction.
The mechanics matter because brain metastases evolve in more than one way. Some shrink after stereotactic radiotherapy. Some remain stable for a time. Others recur at the margins or enlarge subtly before they are obvious on routine review. By making serial change easier to see, the AI-assisted workflow appears to improve the detection of those small transitions, especially when the same lesion is imaged repeatedly across many months of care.
The workflow is as important as the algorithm
This study did not simply compare an algorithm with a human reader. It compared two end-to-end clinical workflows. In the manual process, radiologists performed lesion measurement directly. In the AI-assisted process, the system preselected studies, registered them longitudinally, and generated measurements that could be revised by the reader. That design reflects a real-world radiology environment, where the bottleneck is often not image availability but the time required to assemble and compare the relevant series.
The study also recorded time and the number of mouse clicks for both workflows, though those specific values were not provided in the abstracted material. Even without those metrics, the intended clinical value is clear: fewer manual steps should reduce friction during follow-up interpretation, especially for patients with many lesions and multiple prior scans. The PACS-integrated setting is important because it places the tool where reading already happens, rather than forcing radiologists into an external platform that disrupts routine practice.
The underlying biological rationale is straightforward. Brain metastases are dynamic lesions, and treatment response can vary across metastases within the same patient. A volumetric or semi-automated longitudinal system can help distinguish true progression from measurement noise, postoperative or post-radiation change, and reader variability. That distinction can influence whether a patient continues current local therapy, undergoes closer surveillance, or is evaluated for a new treatment strategy.
What the data support now
This study supports a cautious but tangible conclusion: AI-assisted longitudinal lesion tracking improved follow-up assessment in patients with brain metastases treated with stereotactic radiotherapy and helped identify lesions that later progressed. The study also makes a sharper point about the limits of manual review. Even when expert neuroradiologists agree strongly overall, some lesions still fall through the cracks, and a subset of those lesions later prove to be active disease.
The dataset was modest, with 40 patients and 158 MRIs, so the findings should be read as a proof of workflow value rather than a final validation across all clinical settings. The median Dice coefficient of 0.771 is encouraging, but not definitive for a tool meant to influence oncologic surveillance. Larger studies will need to determine how reliably this approach performs across different scanners, institutions, lesion sizes, and treatment contexts. For now, the clinical message is narrower and more immediate: longitudinal AI support can make brain metastasis surveillance less vulnerable to missed progression, especially when treatment decisions depend on small changes over serial scans.