WHAT IS MEDICAL IMAGE ANNOTATION?
Medical image annotation is the process of identifying, outlining, and labeling anatomical structures, regions of interest, or pathological findings within medical imaging data such as MRI, CT, PET, and ultrasound scans. These annotations form the foundation of reliable datasets used in medical research, AI model training, surgical planning, and clinical imaging workflows involving quantitative output. High-quality annotations require detailed anatomical knowledge, precision, and consistency across imaging volumes.
At Mathelios, the focus is exclusively on expert manual annotation and delineation performed by a trained anatomist. We do not develop AI software or automated segmentation systems. We provide the high-quality human-generated ground truth data those systems depend on, and use our knowledge to improve them.
WHAT DO WE OFFER?
1. MANUAL DELINEATION. An expert traces boundaries structure by structure, slice by slice, with full anatomical understanding of what they are looking at and why it matters, making it the gold standard for research, AI training data, and radiation therapy planning.
Manual delineation of left kidney using a drawing tool (first image) and intensity-based manual delineation of a large left intracerebral hematoma (second image) leveraging Hounsfield unit thresholding. Thresholding can substantially accelerate the delineation of hemorrhagic tissue, given the characteristically high attenuation of acute blood on computed tomography (CT). However, review and manual correction remain essential to fill hypodense gaps, resolve partial volume effects at the boundaries, and exclude adjacent hyperdense structures such as bone or calcifications.
2. 3D SEGMENTATION. Extending delineation into three dimensions, producing voxel-level masks of anatomical structures that enable volumetric measurement, surgical planning, and AI model training across full imaging volumes, allowing monitoring of structural changes over time and visualisation of anatomy in its full spatial context.
Multi-planar visualization of manual left kidney segmentation (axial, coronal, sagittal) with 3D surface rendering, providing a comprehensive view of the delineated volume across all anatomical orientations.
3. HUMAN EXPERT-IN-THE-LOOP. Computer-automated tools can produce a useful first pass, but they still frequently fall short, particularly at boundaries, in pathological tissue, or in atypical anatomy. Expert review and correction ensures the final output meets the highest clinical and research standards. Critically, every correction is also an opportunity: expert-validated annotations can be fed back into the pipeline to continuously retrain and refine the underlying model, turning each human intervention into a long-term investment in automation quality.
Left: Initial automated segmentation of Couinaud liver segments (TotalSegmentator) showing co-segmentation of adjacent gastric tissue at the liver boundary. Center: Close-up view of the segmentation error (arrow). Right: Refined output following manual boundary correction of liver segment 3, with accurate exclusion of the stomach.
Computer-automated brain segmentation on MRI can produce seemingly robust results to the untrained eye, yet on closer inspection, delineations in the frontotemporal and cerebellar regions frequently fall short (see white box, first image), driven largely by partial volume effects and local intensity ambiguities. The regions involved are of central importance in neurodegenerative disease research, yet their segmentations are rarely visually inspected before derived metrics are reported. The same holds for total-body CT segmentation, where under- and oversegmentation and poorly defined boundaries are common, as seen here in the small intestine (purple, second image).
With careful manual refinement, however, the most problematic boundaries can be corrected, bringing the segmentation closer to true anatomy. Perfection is not the goal, reliability is. Left: raw algorithm output. Right: manually refined boundaries, reviewed and improved by an expert anatomist.
WHY GROUND TRUTH MATTERS.
In medical imaging research, ground truth is everything. It is the reference standard against which AI models are trained, validated, and ultimately trusted in clinical practice. And yet, it is only as reliable as the expertise of the person who created it. A mislabeled structure, an imprecise boundary, a missed pathology, these are not small errors. In AI training, they propagate. In research, they skew results. In radiation therapy planning, they have consequences that reach far beyond the dataset.
Radiologists and imaging researchers are (increasingly) aware that most AI models underperform on real-world data compared to published benchmarks, and that behind many of those benchmarks lie annotations that were not always rigorously verified. To move the field forward, we need more reliable high-quality datasets. Even though high-quality "ground truth" takes time and expertise, it is worth every investment of both.
BODY REGIONS
Mathelios specialises in neuro- and musculoskeletal anatomy and pathology, with experience across other organ systems and body regions, including thoracic, abdominal, and pelvic anatomy. More broadly, if it involves (radiological) anatomy, it is worth a conversation. Every project is different, and the scope, modality, and level of detail are always tailored to the question being asked.
Beyond research and clinical applications or surgical planning, Mathelios is equally open to projects in VR/AR development, 3D printing, and any domain where precise anatomical delineation adds value. Whether you are a surgeon, radiologist, a researcher, an engineer, or a developer ... if anatomy is at the heart of your project and you require manual precision, let's talk.