Guidance · Image capture

Medical Image Guidelines

How to capture and upload a tympanic membrane image that OtoSensAI can analyse reliably. A good result depends first on a good image.

Before you capture

1.Equipment

Use a digital otoscope, or an endoscope fitted with a camera capable of clear, focused close-up images of the ear canal and eardrum. A steady light source is essential — most classification errors trace back to poor lighting or blur.

Clean the speculum or tip before each use.
Choose the largest speculum that fits the canal comfortably — it gives a wider, steadier view.
Make sure the lens is clean and free of smudges or condensation.

2.Prepare the ear

If wax (cerumen) or debris obscures the eardrum, it should be safely cleared by a trained person before imaging — the AI cannot see through an obstruction.
Straighten the ear canal: gently pull the pinna up and back in adults, and down and back in young children.
Keep the patient still; brace your hand against the head to reduce motion blur.

What a good image looks like

✓ Aim for

The eardrum centred and filling most of the frame
Sharp focus — landmarks such as the malleus handle and light reflex visible
Even, natural lighting across the drum
Canal walls minimal; the drum clearly the subject
One clear single view per upload

× Avoid

Blur from movement or wrong focus distance
Glare or over-exposed white "hot spots"
Images that are too dark or underexposed
Wax, hair, or debris covering the drum
Off-centre shots showing mostly canal wall

A quick self-check before uploading

Can you clearly see the eardrum — in focus, well-lit, and unobstructed? If not, retake the image rather than uploading a borderline one. A borderline image produces a borderline output, which invites a decision it cannot support.

3.File requirements

Requirement	Specification	Enforcement
Accepted formats	JPG, PNG, or WEBP	Server-side — non-conforming uploads are rejected
Maximum file size	10 MB per image	Server-side
Orientation	Upright; do not stretch or distort	Guidance
Editing	No filters, text overlays, or annotations on the image	Guidance
Per upload	One tympanic membrane view	Server-side
Image metadata (EXIF)	Stripped automatically at upload, before storage	Automatic — handled by us
Filename	Discarded and replaced with a generated identifier	Automatic — handled by us

4.Privacy

No personal information may appear in the image, its filename, or any accompanying note. Responsibility is divided according to what each party can control.

Task	Responsibility	Why
Strip EXIF and all image metadata	Voxmedai — automatic	You cannot inspect it; we can remove it
Discard the original filename	Voxmedai — automatic	Handled at ingestion, before storage
Assign a pseudonymous identifier	Voxmedai — automatic	Handled at ingestion, before storage
Frame the shot to exclude faces, tattoos, jewellery, and identifying marks	You	Only the person holding the scope controls the frame
Keep names, dates of birth, IDs, and contact details out of the notes field	You	Only you know what you are typing
Obtain consent before capturing a patient's image	You	Only you are present with the patient

Uploaded images are de-identified at the point of upload and, where you have consented, used to improve the AI model. Keeping identifying data out of the frame and out of the notes field protects everyone.

5.Consent

If you are a clinician uploading a patient's image, ensure you have obtained appropriate consent for AI-assisted analysis, consistent with the Terms & Conditions. If you are uploading your own image, you are consenting on your own behalf.

Images of children

Under Section 9 of the DPDPA, processing a child's personal data requires verifiable parental consent.

Where an image is of a person under 18, the parent or legal guardian must complete our verifiable parental consent process before the image is accepted. A self-declaration is not sufficient.

Images of minors are not accepted through the Service until that process is completed.

6.What a good image cannot fix

These guidelines exist because image quality is the largest controllable source of classification error. Following them well improves the result — but it does not make the result reliable.

A technically excellent image of pathology the model was never trained on will still produce a confident, wrong answer.
The model sees one frame. You see the patient, the history, the other ear, and the preceding weeks.
Nothing in this document changes the Medical Disclaimer, or the requirement that every result be verified by a qualified clinician before it informs any decision.