DicomDoodle: Human-in-the-Loop Annotation of 3D MRI DICOM Images with AI Support and DICOM-SEG Compatibility.
Authors
Affiliations (12)
Affiliations (12)
- Department of Computational and Data Sciences, Indian Institute of Science, Bengaluru, Karnataka, India.
- Department of Neurosurgery, All India Institute of Medical Sciences, New Delhi, India.
- Neuropathology Laboratory, All India Institute of Medical Sciences, New Delhi, India.
- Department of Neuroimaging and Interventional Neuroradiology, All India Institute of Medical Science, New Delhi, India.
- Department of Neurosurgery, Postgraduate Institute of Medical Education and Research (PGIMER), Chandigarh, India.
- Department of Neurosurgery, King George's Medical University (KGMU), Lucknow, India.
- Department of Neurosurgery, National Institute of Mental Health and Neurosciences (NIMHANS), Bengaluru, India.
- Department of Neurosurgery, All India Institute of Medical Sciences (AIIMS), Bhubaneswar, Odisha, India.
- Department of Neurosurgery, Sanjay Gandhi Postgraduate Institute of Medical Sciences (SGPGIMS), Lucknow, India.
- Department of Neurosurgery, Sree Chitra Tirunal Institute for Medical Sciences and Technology (SCTIMST), Thiruvananthapuram, Kerala, India.
- ARTPARK-IISc, Bengaluru, India.
- Department of Computational and Data Sciences, Indian Institute of Science, Bengaluru, Karnataka, India. [email protected].
Abstract
Accurate medical image annotation requires curation and remains a major bottleneck due to its time-intensive nature and inter-observer variability. To address this, we developed DicomDoodle, an open-source, Python-based tool for human-in-the-loop annotation that integrates a fine-tuned YOLO11 detector with the Segment Anything Model (SAM) 2.1 for prompt-based segmentation, alongside manual annotation and correction tools. DicomDoodle saves annotations as DICOM Segmentation (DICOM-SEG) objects, which passed automated structural conformance testing (22 checks across all 122 series) and were successfully imported and rendered in OHIF Viewer. Annotation performance was evaluated using brain MRIs from 37 patients (122 series) across three Indian hospitals, with the patient as the unit of analysis. SAM 2.1 and SAM 3 performed best among the six benchmarked SAM and MEDSAM models. Patient-level mean slice and volume IoU were poor for the fully automated workflow, at 0.109 (95% confidence interval 0.087-0.130) and 0.133 (0.102-0.164), respectively, revealing its limitations. It improved to 0.388 (0.329-0.446) and 0.338 (0.280-0.396) with slice-range prompting, and 0.858 (0.841-0.875) and 0.865 (0.843-0.886) with per-slice bounding-box prompting. For complete, unfiltered volumes, the automated workflow detected 95% of slices containing lesions. Still, it labeled about 90% of lesion-free slices as lesion-present; prompting removed almost all of these false positives. Segmentation accuracy did not degrade with lesion extent. In routine use, annotation sessions were provisionally estimated as shorter than a fixed manual reference. DicomDoodle offers an accessible, open-source platform that combines AI-assisted prompting with native DICOM-SEG generation to assist annotation.