Medical Image Annotation Pricing: What Radiology and Pathology Labelling Costs

Why Medical Annotation Costs More Than General Computer Vision

In standard computer vision annotation — bounding boxes around cars, pedestrians, or retail products — the dominant cost is annotator labour at commodity rates. A competent annotator can box 200–500 objects per hour once calibrated on a task. Medical image annotation breaks this model entirely.

Medical annotations require clinical expertise. A radiologist identifying a pulmonary nodule on CT or a pathologist grading Gleason score on a prostate slide is providing specialist medical judgement, not pattern-matching on a label schema. According to the Global Market Insights 2025 Medical Image Analysis report, the medical imaging AI market is projected to reach USD $21 billion by 2030 — and the bottleneck to growth is consistently cited as high-quality annotated training data, not modelling capability.

This creates a structural pricing difference: a radiologist reviewing 15–25 cases per hour at AUD $180–$250/hour as part of an annotation workflow produces a per-image cost floor that has nothing to do with platform licensing. It is a clinical labour cost that no software optimisation eliminates entirely.

AI-assisted pre-labelling (where a model proposes labels and a clinician reviews and corrects) can reduce annotation time per image by 40–65% on well-defined tasks such as organ contouring in CT. But it does not eliminate the clinician review step — for FDA submission pathways, every label in the training set requires human sign-off and audit trail documentation. Pre-labelling reduces cost; it does not remove the cost floor set by clinical review.

Radiology Annotation Pricing: Cost Drivers by Modality and Task

Radiology encompasses CT, MRI, X-ray, ultrasound, PET-CT, and fluoroscopy. Costs vary significantly across these modalities because of image complexity, slice count, and the level of clinical expertise required for the annotation task.

Chest X-ray annotation

Chest X-rays are the highest-volume radiology AI training task globally. For presence/absence classification (pneumonia, pleural effusion, pneumothorax) with radiographer review and radiologist QA, typical costs range from AUD $4–$12 per image at scale. Multi-label classification adding precise anatomical localisation (bounding box around pathological finding) runs AUD $10–$25 per image. The CheXpert dataset (Stanford, 2019) demonstrated that inter-radiologist agreement on chest X-ray labels averages κ = 0.75–0.82 on common pathologies — which is why gold-standard adjudication protocols add cost but are not optional for production AI.

CT scan annotation

CT studies are multi-slice (typically 80–500+ axial slices per study), which makes per-image pricing less meaningful than per-study pricing. Organ segmentation tasks (liver, spleen, kidneys in abdominal CT) with semi-automatic tools and radiologist review run AUD $30–$90 per study. Complex oncology annotation — tumour segmentation with lesion-level attributes across multi-phase CT — can reach AUD $80–$200 per study depending on lesion count and slice depth required. Our CT scan annotation service covers these workflows with radiologist-led review and DICOM SR output.

MRI annotation

MRI annotation for brain AI — lesion segmentation in multiple sclerosis, tumour delineation in glioma, anatomical landmarking for neurodegenerative disease — is among the most expensive imaging annotation tasks. Multi-sequence consistency (T1, T2, FLAIR must be co-registered and labelled consistently) and neuroradiologist review requirements push per-study costs to AUD $60–$250 for complex protocols. Cardiac MRI segmentation (LV/RV contouring across frames) adds temporal consistency requirements that further increase review time.

Ultrasound annotation

Ultrasound AI training data is lower-cost than CT or MRI because individual frames are 2D and real-time acquisition produces large volumes quickly. Foetal biometry tasks (head circumference, biparietal diameter measurement) with sonographer review run AUD $3–$8 per frame. Thyroid nodule classification (TIRADS) with radiologist review is AUD $8–$20 per image. The challenge with ultrasound is operator-dependent image quality variation, which increases QA overhead at the dataset level.

Pathology Annotation Pricing: Why Whole-Slide Images Are Expensive

Histopathology annotation is the highest-cost segment of medical image annotation. A single whole-slide image (WSI) scanned at 40× magnification produces a file of 2–8 GB and an image that would span approximately 100,000 × 80,000 pixels at full resolution. Annotating this at clinically meaningful granularity — identifying tumour regions, classifying cell types, grading Gleason or Nottingham scores — requires a board-certified pathologist working within a specialised WSI viewer (Qupath, ASAP, Aperio ImageScope, or a custom DICOM viewer with multi-resolution tile rendering).

According to a 2024 Nature Medicine review on computational pathology, inter-pathologist agreement on tumour boundary delineation averages κ = 0.61–0.78 for common carcinomas, falling to κ = 0.45–0.58 for rare or ambiguous cases. This agreement ceiling means multi-pathologist adjudication is not a quality enhancement — it is a minimum requirement for production oncology AI. Two-pathologist adjudication with a third-party arbitration protocol for disagreements typically doubles the annotation labour cost per slide versus single-pathologist review.

Realistic pathology annotation pricing by task:

• Tumour presence classification (slide-level): AUD $25–$60 per slide with single pathologist
• Tumour region delineation (ROI annotation): AUD $50–$150 per slide with single pathologist
• Cell-level annotation (nuclei, mitotic figures): AUD $80–$300+ per slide depending on cell density and label schema
• Gleason / Nottingham grading with adjudication: AUD $80–$250+ per slide with two-pathologist protocol
• IHC biomarker scoring (ER, PR, HER2, Ki-67): AUD $40–$120 per slide, depending on scoring granularity

Our histopathology annotation service provides board-certified pathologist annotation with multi-reader adjudication and full WSI tooling — contact us to discuss slide complexity and throughput requirements.

Regulatory Overhead: What FDA, HIPAA, and TGA Add to Medical Annotation Cost

Medical AI projects targeting regulatory submission carry compliance costs that general computer vision annotation does not. These costs are real and predictable — they should be budgeted explicitly, not treated as unknowns.

FDA 21 CFR Part 11 requires electronic records to include audit trails, timestamped annotator actions, access controls with individual authentication, and the ability to reconstruct the state of any annotation at any point in time. For annotation vendors, this means purpose-built provenance infrastructure — not a standard off-the-shelf annotation platform. Implementing 21 CFR Part 11-aligned logging adds approximately 20–30% to annotation infrastructure cost, which flows through to per-image pricing for FDA-pathway projects.

HIPAA Business Associate Agreements (BAAs) are required for any vendor handling US patient data. BAA execution, data processing agreement review, and the security controls they require (encryption at rest and in transit, access logging, breach notification procedures) add administrative and technical overhead that smaller annotation vendors cannot absorb. Verify BAA capability before sharing any DICOM data containing patient information.

For Australian medical AI projects, TGA requirements under the Software as a Medical Device (SaMD) framework align closely with FDA guidance on AI/ML-based SaMD (the 2021 FDA action plan). While TGA does not require 21 CFR Part 11 specifically, the equivalent evidence requirements for training data provenance are comparable. Budget 15–25% for compliance overhead on TGA-pathway projects.

De-identification: The Step Before Annotation That Most Teams Underbudget

DICOM files contain patient-identifiable information in header fields (patient name, DOB, hospital ID, referring physician) and sometimes in image-burned overlays. Before any annotation can begin on patient data, de-identification must be performed and verified. This step is consistently underbudgeted because teams assume it is automated — and often it is not fully automatable.

Standard DICOM de-identification tools (DicomCleaner, pydicom, Orthanc DICOM server) handle header field removal. Burned-in annotations — patient names or dates overlaid on the image itself — require visual review and image-level processing. For large datasets, this can add AUD $0.50–$3 per study in verified de-identification cost before the annotation clock starts.

Factor de-identification into your annotation budget from the start. It is not optional for regulatory pathways and is a precondition for sharing data with any external annotation vendor under HIPAA or the Australian Privacy Act.

How to Get an Accurate Medical Annotation Quote

Medical annotation quotes are frequently inaccurate at project start because vendors are quoting blind on task complexity. The inputs that make a quote accurate are:

• Modality and scan protocol (CT with/without contrast, 1.5T/3T MRI, 40× vs 20× histology scan)
• Annotation task schema: label classes, label types (bounding box, polygon, mask, keypoint, classification), and attributes per label
• Sample images: even 10–20 representative cases let a vendor estimate annotation time per case accurately
• Annotator credential requirement: radiographer vs radiologist vs sub-specialist; pathologist board-certification level
• Compliance pathway: HIPAA BAA required, FDA 21 CFR Part 11 provenance, TGA SaMD documentation
• Volume and timeline: cases per week, total dataset size, phased delivery schedule
• Output format: DICOM SR, nnUNet NIfTI, COCO JSON, custom schema

The single most effective step before committing to a volume order is a structured free pilot: 20–50 images annotated to your schema, reviewed against your gold standard, with unit cost and throughput measured. This eliminates the most common source of medical annotation project failure — scope misalignment between what the buyer thought they were getting and what the vendor thought they were delivering.

Our radiology annotation service and histopathology annotation service both offer structured free pilots — contact us with your modality, task schema, and compliance requirements and we will scope the pilot against your specific data.

Reducing Medical Annotation Cost Without Compromising Quality

The most effective lever for reducing medical annotation cost is AI-assisted pre-labelling. For tasks with well-trained foundation models — organ segmentation in CT, nuclei detection in H&E histology, optic disc segmentation in fundus photography — an AI pre-labelling step can generate initial labels that a clinician then reviews and corrects. Published studies on model-assisted pathology annotation (Tizhoosh et al., 2021; Srinidhi et al., 2021) report 40–60% reductions in pathologist annotation time per slide when pre-labelling quality is sufficient.

The key constraint is pre-label quality thresholds. If the AI model produces poor initial labels on difficult cases, review-and-correct takes longer than annotation from scratch. The best practice is to stratify: auto-accept high-confidence AI labels with light human QA spot-check, route moderate-confidence cases to radiographer/technician correction, and route low-confidence or complex cases to full specialist review. This three-tier approach achieves the best cost reduction without concentrating specialist clinician time on cases that do not need it.

Schema discipline is the other major cost lever. Adding label classes or attributes after annotation begins requires relabelling of the entire dataset — the most common source of cost overrun in medical annotation projects. Defining a minimal viable schema from the start, validated against your model architecture requirements before a single image is annotated, prevents this. Related reading: how to write annotation guidelines that don't need constant revision.

For teams working with radiological data and unsure about annotation costs or scope, our pricing page provides general guidance, and a free consultation can help scope your specific project before you commit to any volume. See also our overview of data annotation cost and pricing for non-medical modalities.

Frequently Asked Questions