Segmentation accuracy
Our AI models deliver reliable results across studies. Even under strict quality criteria (requiring accurate segmentation in at least 6 out of 7 organs), failure rates were below 4% when tested on a hard and diverse data set of 259 rats and mice using cross-validation. This means researchers can trust the segmentations to be consistent, reproducible, and ready for dosimetry.
Accuracy and consistensy is of the utmost importance for automated segmentation of CT images. At preclinIQ we thoroughly test our automated segmentations against computer aided annotations and manual annotations. This is done to ensure both human-like precicion and machine-like consistensy.
Our models are stress-tested on hundreds of scans across multiple years, scanners, and study protocols, covering both mice and rats. This broad validation guarantees that performance is not constrained to a single dataset but generalizes to the diverse imaging conditions found in real research environments.
Key highlights
This one-vs-all comparison shows how much each expert agrees on average with the others. Our AI matches this level of agreement at it is performing on par with the three expert annotators.
- Proven reliability: performance on par with human annotators
- Validated for PET/SPECT dosimetry: errors typically below 10% in key organs
- Continuous improvement of segmentation: the segmentations are continously improved by increasing dataset volume, quality and diversity
- Post-processing: All organ segmentations are processed to fit the expected number of ROIs.
- Built-in safeguards: automatic QC flags outliers for optional manual inspection
Read more about the automation and time saved.
Does it fail?
No AI system is perfect, but at preclinIQ every scan is checked automatically before results are released.
We validate each segmentation using organ morphology—making sure that the right number of organs are detected and that their shape, size, and relative position make biological sense. If something looks unusual, the system flags it for optional manual review.
Segmentation quality also depends on imaging quality: scans with good contrast and minimal artifacts lead to the most precise results. Even so, our models are designed to handle a wide range of imaging conditions, and the built-in safeguards ensure that unreliable outputs never go unnoticed.