The fundamental challenge in image annotation is the speed-quality tradeoff. Manual annotation of complex segmentation masks is slow and expensive. Automated annotation is fast but introduces systematic errors especially at object boundaries, occluded regions, and in domain-specific images (medical scans, satellite imagery, unusual lighting conditions) that differ from the model's training distribution.
Our approach combines SAM2 (Meta's Segment Anything Model 2) for AI-powered pre-annotation with expert human validation. SAM2 pre-draws segmentation masks or bounding box suggestions at 40–60% of the speed of manual annotation. Human annotators then validate, correct, and refine these suggestions with particular attention to boundary accuracy, small objects, occluded targets, and edge cases that AI models consistently miss.
For medical imaging specifically (DICOM radiology, pathology slides, ultrasound), we maintain a separate annotator pool of MBBS-qualified clinicians and radiologists. Medical image annotation requires clinical knowledge the difference between a tumour margin and a normal tissue boundary is not something that can be specified in annotation guidelines alone; it requires annotators who have studied anatomy and pathology.
