Laurens Topff, MD, PhD, Radiologist and Medical Information Officer for AI at the Netherlands Cancer Institute in Amsterdam, combines expertise in diagnostic radiology and imaging informatics to guide the development and deployment of AI applications in healthcare. His work focuses on translating digital innovations into meaningful improvements in real-world clinical practice. Topff presented his work at the SRG Pillar meeting for the Center for Intelligent Imaging (ci2) at the University of California, San Francisco. His presentation, "Bridging the Gap to Clinical Practice: Building Deployable Brain Metastasis AI," showcased his project.
Although there are many use cases that provide value, in most clinical settings, AI research does not lead to a clinical impact; this phenomenon is called the 'AI Gap.' The translational gap stems from several factors, including non-reproducible or underperforming results. "Bridging the gap requires a multidisciplinary approach," Topff said.
Barriers to the widespread application of commercially available AI tools include errors in detection, segmentation, and registration, as well as usability issues.
After seeing an increasing number of brain metastasis patients at his cancer institute, leading to more MRI scans and lengthy reports, his team built a clinical classification model that was trustworthy and clinically feasible. Previously, AI tools for brain metastasis on MRI were not clinically viable. To create a trustworthy and deployable AI, Topff and his team needed a tool with high detection sensitivity for even the smallest lesions, an 'acceptable' number of false positives (one per scan), reliable segmentation for volumetric analysis, and automatic registration for longitudinal follow-ups. The AI tool also needed to be generalizable across the vast heterogeneity of scanners and institutions and user-friendly, with an easy clinical workflow integration.
Along with his team's focus on building a better model, the need for a data-focused approach was evident. "Improving data quality is seen as a more important driver for clinical trustworthiness and generalizability," Topff said. "The data-centric parts ... can really drive adoption and clinical use of these tools."
While evaluating usable datasets, Topff noted the UCSF-BMSR as the most useful for his team, specifically the BraTS version.
In research conducted in early 2025, Topff and his team compiled a multicenter dataset comprising 1,985 scans from 30 scanners. They fed this information, along with annotation instructions, into their AI model, BrainMets. The model achieved high sensitivity, produced few false negatives, and demonstrated reliable segmentation and registration. Compared with baseline models, BrainMets was very successful.
To learn more about the upcoming SRG Pillar meetings, visit the ci2 events page.
