Diagnostic AI runs up less against the limits of computing power and more against a shortage of high-quality annotated data. Physician and co-founder of the startup Medano, Pavol Praženica, describes how manual dataset preparation slows model development and how it can be sped up. He presents a platform that integrates image collection, annotation, and model training, and shows early clinical benefits from 3D surgical planning to supporting prostate screening.
The biggest obstacle: data, not computing power
Reference datasets are key for training diagnostic models, but they are created through time-consuming manual annotation. Radiologists in hospitals label structure by structure, which for X-ray images means roughly a thousand patient images per diagnosis. Such a process is hard to scale and ties up the capacity of specialists, who are already in short supply.
For CT and MRI the situation is even more demanding, because many diagnoses span dozens of slices. A renal carcinoma may be visible on roughly fifty slices, and if three teams of radiologists are to prepare at least a thousand reference cases, the work takes years. In practice, according to the Medano team, it would exceed four years of continuous effort.
One platform for the entire annotation cycle
Medano has therefore developed a web platform that runs in the cloud and communicates directly with the hospital PACS system. Based on user instructions, it automatically downloads medical images and offers its own tools for faster labeling. The interface is designed so that even physicians without prior experience in dataset preparation can handle it.
The workflow is simple: a clinician first manually annotates approximately thirty cases, from which an initial pre-annotation model is trained. Every subsequent human adjustment refines the model, speeding up and streamlining the entire cycle. The company places an emphasis on user experience and integrates steps that are commonly done in three or more software tools into a single environment.
From 3D visualizations to prostate screening
Surgeons in Belgium use the models for automatic 3D visualization of renal vasculature and kidney carcinoma during preoperative preparation. Patient-specific visualizations are available immediately before the procedure. Operations then take less time, there are fewer complications, and recovery is shorter.
The team also focused on screening for prostate cancer, which ranks among the most common malignant tumors in men, where waiting times have shifted from weeks to months. The model, trained on a combination of MRI images and histopathology, automatically reviews pelvic MRI, estimates the presence of a tumor, and assigns a PI-RADS score. In radical prostatectomy, additionally, in collaboration with the manufacturer of the portable CT scanner XOS, they created a workflow in which a 3D model is generated from MRI before the operation and, after the prostate is removed, the specimen is scanned with a portable CT. By comparing the preoperative and postoperative 3D visualizations, the model helps verify whether the tumor was removed completely and, if necessary, guide the surgeon to extend the procedure.
