AI Acne Detection: How Skin Apps Find Acne in a Photo and Why They Miss Some Spots

AI acne detection works by finding small candidate spots in a face photo, drawing a box around each one, and sorting it into a class such as blackhead or whitehead, papule or pustule, nodule, or a leftover mark. Apps miss spots that are tiny, close to skin color, hidden on the jawline, or flattened by poor light, and they often confuse healed marks and scars with active pimples.

Below is how that pipeline works step by step, what published smartphone studies say about accuracy, where the errors come from, and how to take photos that give a detector a fair chance. If you want the grading side of the story, read our companion piece on acne severity grading.

How does an app find acne in a photo?

Most acne detectors run an object-detection model that scans the photo for lesions and labels each one, then counts them by type. A 2022 study in Diagnostics describes a typical setup: a Faster R-CNN detector trained on 1,572 smartphone selfies containing 41,859 labeled lesions, sorting each box into blackheads/whiteheads, papules/pustules, nodules/cysts or acne scars, with a second model turning the counts into a severity grade.

In plain steps, the pipeline usually looks like this:

1. Quality check. The app rejects photos that are too dark, too far away or blurred. In the study above, the app checked that the face was about 20 cm from the camera and that lighting was sufficient before it would capture.
2. Face and zone finding. Landmarks locate the forehead, cheeks, nose, chin and jaw so spots can be counted per region and hair, eyes and lips are ignored.
3. Lesion detection. The model proposes boxes around anything that looks like a spot and gives each a confidence score.
4. Classification. Each box gets a label. Some systems also keep a separate class for marks left after a pimple heals, so those aren't counted as active acne.
5. Grading. Counts per type and zone become a mild, moderate or severe label, often modeled on dermatologist scales such as one Japanese standard that counts inflamed papules and pustules on half the face, calling 0 to 5 mild and 6 to 20 moderate.

Other systems skip whole-face detection. One app described in Applied Intelligence asks users to circle lesion areas by hand and upload them, then classifies the crops into normal skin, comedone, papule, pustule or nodule, because the authors dropped the face-segmentation network to keep the model small enough for a phone.

What types of acne can a detector tell apart?

A detector can only tell apart the classes it was trained on, and those usually follow the standard clinical list. DermNet describes acne as a mix of open and closed comedones (blackheads and whiteheads), inflamed papules and pustules, and in severe cases nodules and pseudocysts, plus the red or brown marks and scars that can follow. If you want to read the spots yourself, our guide to papules vs pustules walks through each one.

The study numbers back up the "easy" and "hard" columns. In the Diagnostics model, average precision was 0.68 for nodules/cysts and 0.64 for papules/pustules, but only 0.44 for scars and 0.40 for blackheads/whiteheads. The authors explained that most scars are small and differ little in color from the skin around them, while nodules were easier because they are larger and usually red.

How accurate is AI acne detection?

Published smartphone systems agree with dermatologists about as often as dermatologists agree with each other, which is less often than most people expect. A study funded by La Roche-Posay trained an algorithm on 5,972 smartphone images from 1,072 patients in France, South Africa, China and India, each with a front view and left and right profiles. Its final version reported F1 scores of 84% for inflammatory lesions, 61% for non-inflammatory lesions and 72% for post-inflammatory hyperpigmentation, and agreed with dermatologists on the overall grade 68% of the time.

The same paper is a good reminder that the "ground truth" is soft. When dermatologists graded the clinical test set, they agreed with each other 59% of the time from images and 67% face to face. The other studies show the same pattern: the hand-circled app above reported 93.19% grading accuracy from images (94.56% when patient details were added) and was judged noninferior to each of three dermatologists, but its testing used mostly Chinese participants from one hospital and front cameras did significantly worse than rear cameras for both the app and the clinicians.

So treat an acne count as an estimate with a margin of a few spots either way, most useful for spotting a trend across weeks.

Why do skin apps confuse acne with marks, scars and freckles?

A single photo shows color and shape, not history or depth, so anything brown, red or bumpy can look like a pimple to a model. Four confusions come up again and again:

• Healed marks. Flat brown patches that range from light brown to black and sit where an earlier skin problem healed look like a fading spot. DermNet notes they are more common and more intense in darker skin, and can last longer there, which raises the false-positive rate for exactly the people most likely to have them. Red marks are similar; our guide to red marks after acne explains why they linger.
• Scars. Shallow pits read as shadows, and a raised scar can look like a papule head-on.
• Freckles and moles. A small, round brown dot gives a model little to separate it from a healing spot, especially at selfie resolution.
• Redness that isn't acne. Flushing, irritation or a rash spreads color across a zone and can make the model see clustered papules.

Even dermatologists find the mark category hard. In the La Roche-Posay study, dermatologists counting post-inflammatory hyperpigmentation face to face reached an agreement score of only 0.16 to 0.55, which the authors rated poor to fair. A model trained on those labels can't be much more consistent than they are.

Why do apps miss spots on the jawline and cheeks?

Apps miss jawline and side-cheek spots mainly because a front-facing selfie shows those areas at a steep angle, small and foreshortened. The jaw curves away from the lens, so a pustule near the ear may cover a few pixels, and the lower jaw often falls into shadow from the chin. That's why research datasets take more than one view: both the La Roche-Posay study and the Diagnostics study used a front photo plus left and right profiles.

The other common reasons for misses:

• Light from above. Ceiling lights flatten small bumps on the cheeks and throw shadow under the jaw.
• Distance. Holding the phone at arm's length shrinks every spot. Smartphone images are already lower in quality than dermatology camera images, so pixels matter.
• Front camera. In the Applied Intelligence study, front cameras performed significantly worse than rear cameras.
• Makeup, tinted SPF and filters. Anything that evens tone hides the color contrast a detector depends on.
• Hair and beards. One study simply excluded patches containing hair, so spots under a fringe or stubble may never be counted.

How do I take photos that catch every spot?

Take three photos, front, left and right, in soft daylight from the same distance each time. The American Academy of Dermatology's photo advice for patients is a good base: use bright natural light, such as from a window, avoid shadows and glare, and take more than one picture of each view. Our guide to the best lighting for skin photos goes deeper.

1. Wash your face, skip makeup and tinted sunscreen, and pull hair back from the forehead and jaw.
2. Face a window with daylight on your face, not behind you. Turn off overhead lights if you can.
3. Take the front view at eye level, phone roughly 30 to 40 cm away, chin level.
4. Turn your head about 45 degrees left, then right, so each cheek and the jawline face the lens. Keep the light on the side you're photographing.
5. For a hard spot, add a close-up. Use the rear camera if someone can help, since it usually has the better sensor.
6. Repeat at the same time of day, in the same spot, every one to two weeks.

How can a selfie skin scan help with acne?

A regular scan is most useful as a consistent record, because a few weeks is how long acne treatments need. The AAD says to give a treatment at least 4 weeks, expect some improvement in 4 to 6 weeks, and allow two to three months or longer for clearing, which is long enough for memory to blur.

Pori, our iPhone app, takes a selfie and returns an overall skin score from 0 to 100 with a breakdown that includes Pores & Texture and Pigmentation & Tone, read region by region across a 468-point face mesh. It doesn't give a lesion count; what helps with acne is the before-and-after comparison, progress charts and a skin diary for sleep, stress and diet, so you can see whether texture and marks are trending the right way. The first scan is free; Pro costs $9.99 a week or $49.99 a year in the US as of October 2026. For a full method, see how to track acne progress.

An app is not a diagnosis. See a dermatologist for painful lumps under the skin, new scarring, acne that hasn't improved after two to three months of steady treatment, or any spot that bleeds, grows or changes.