Smartphone Scan Classification for Remote Orthodontic Monitoring Across a 24-Clinic Group
Let's Connect
Overview
What we built
A 24-clinic orthodontic group had clinical coordinators manually reviewing every weekly smartphone scan patients submitted between visits, and the backlog was growing by thousands of photos a week. We built an automated triage pipeline that classifies each scan and routes only the ones that need a person.
In plain terms: patients on aligner treatment were asked to send in a smartphone scan of their teeth every week between in-person visits, and every single one of those scans landed on a clinical coordinator's desk for manual review. With thousands of photos arriving weekly, coordinators fell behind, triage decisions varied from one clinic to another, and patients were left waiting days just to hear whether their treatment was still on track.
We built an image recognition pipeline that classifies each scan for photo quality, aligner seating and tracking status the moment it arrives. Unusable images trigger an automatic retake request, routine on-track scans are auto-acknowledged with templated guidance, and only the cases that actually need a person are routed to clinicians with side-by-side scan history for rapid review. Median review turnaround fell from 3 days to under 8 hours across all 24 clinics, 72% of routine scans are now auto-acknowledged without coordinator involvement, unusable submissions dropped from 19% to 6%, and in-person progress visits per patient fell by roughly 25% over a 12-month comparison period.
The Problem
Weekly scans reviewed manually
Across the group's 24 clinics in the US and Canada, remote monitoring ran on a simple idea: ask patients to submit a weekly smartphone scan of their teeth, and have a clinical coordinator check it between visits. Every single scan, regardless of whether it showed anything worth flagging, went through the same manual review.
That approach did not scale with the volume it generated. Thousands of photos arrived every week across the group, and coordinators working through them by hand produced triage decisions that varied noticeably from one clinic to another, so a scan that got flagged in one clinic might be waved through in another.
Patients felt the backlog directly. Waiting days to hear whether their aligner treatment was on track was the norm rather than the exception, which undercut the whole point of weekly check-ins: catching a tracking issue early enough to act on it.
Every scan reviewed by hand
Clinical coordinators manually reviewed every weekly smartphone scan submitted by patients, regardless of whether the image showed anything worth flagging.
Thousands of scans weekly
The group received thousands of photos every week across its 24 clinics, a volume manual review could not keep pace with.
Inconsistent triage
Triage decisions varied noticeably between clinics, so similar scans could be handled differently depending on which clinic reviewed them.
Days-long patient wait
Patients waited days to hear whether their aligner treatment was on track, undercutting the purpose of weekly remote check-ins.
What it was costing them
Every scan waiting in the backlog was a patient wondering whether their treatment was still on track, and days-long waits weakened the whole reason for asking for weekly scans in the first place. Inconsistent triage between clinics meant the group could not point to one standard of remote monitoring, and coordinators working through thousands of photos a week had no way to focus on the scans that actually needed attention.
The Solution
Automated scan triage pipeline
We built an image recognition pipeline that classifies each smartphone scan the moment it arrives, checking photo quality, aligner seating and tracking status before a coordinator ever sees it. Scans that are unusable, blurry, poorly lit or badly framed trigger an automatic retake request straight back to the patient, rather than sitting in a queue for a coordinator to reject later.
Everything else is sorted into review queues by urgency instead of arriving in one undifferentiated stream. Routine on-track scans are auto-acknowledged with templated guidance, so patients whose treatment is progressing normally get a fast, consistent response without a coordinator needing to look at every image.
Only the scans that actually need a person are routed to clinicians, and they arrive with side-by-side scan history attached, so a clinician reviewing a flagged case can see how the patient's tracking has changed over recent weeks rather than judging one photo in isolation.
Key decisions
Classify before routing
Every scan is classified for photo quality, aligner seating and tracking status on arrival, so routing decisions are made before a coordinator gets involved.
Reject unusable images automatically
Unusable scans trigger an automatic retake request at capture time, catching quality problems before they ever reach a review queue.
Auto-acknowledge routine cases
Routine on-track scans are auto-acknowledged with templated guidance, freeing coordinators from reviewing cases where nothing needs to change.
Sort by urgency, not arrival order
Scans that need attention are sorted into review queues by urgency rather than handled in the order they happened to arrive.
Give clinicians scan history
Flagged cases route to clinicians with side-by-side scan history, so a review considers how tracking has changed rather than one photo alone.
Measurable Impact
What changed after launch
The backlog stopped being the bottleneck. Median scan review turnaround fell from 3 days to under 8 hours across all 24 clinics, and 72% of routine on-track scans are now auto-acknowledged without coordinator involvement, freeing that time for the scans that actually need a clinician's judgement.
Quality improved at the source too. Unusable scan submissions fell from 19% to 6% once automated retake prompts caught problems at capture time, and with monitoring working as intended, in-person progress visits per patient dropped by roughly 25% over a 12-month comparison period.
Review turnaround
3 days across the group
Under 8 hours across all 24 clinics
Routine scan handling
Every scan reviewed manually
72% auto-acknowledged without a coordinator
Unusable submissions
19% of scans
Down to 6% with automated retake prompts
In-person visits
Baseline visit frequency per patient
Down roughly 25% over 12 months
Headline results
Median scan review turnaround cut from 3 days to under 8 hours across all 24 clinics
72% of routine on-track scans auto-acknowledged without coordinator involvement within 4 months of rollout
Unusable scan submissions fell from 19% to 6% after automated retake prompts at capture time
In-person progress visits per patient reduced by roughly 25% over a 12-month comparison period
Tech & Tools Used
What powered the build
Every tool below earned its place in this engagement. Here is the part each one played.
Python (FastAPI)
Serves the classification API that receives each smartphone scan and returns photo quality, seating and tracking classifications.
PyTorch
Trains the classification models that judge photo quality, aligner seating and tracking status from each submitted scan.
ONNX Runtime
Runs the exported classification models in production, keeping classification fast enough that patients get a near-immediate retake request when needed.
React Native
Powers the patient-facing app where scans are captured and submitted, and where retake requests and templated guidance are delivered.
Node.js
Runs the services that route classified scans into review queues by urgency and connect the app, the classification pipeline and clinician tools.
PostgreSQL
Stores scan classifications, review-queue assignments and each patient's scan history for clinicians to reference.
Redis
Queues and caches incoming scans so classification keeps pace with the volume arriving from all 24 clinics.
AWS S3 + CloudFront
Stores submitted scan images and delivers scan history quickly to clinicians reviewing flagged cases.
Docker
Packages the classification pipeline and its dependencies into consistent containers deployed across the group.
Ready to Build your Dental & Orthodontics Business with Image Recognition & Classification
Ask Byte
Ask Byte
Typically replies instantly
just Now
Hi! I'm OrganByte's assistant. How can I help you today?
AI-generated content may be incorrect
