THE EXPERIMENT
I wanted to see if a model could do something a little more useful than saying “melanoma” or “not melanoma.”
So I trained one to point.
Most of the first ideas I came across treated the problem as classification. I wanted to try something slightly different: can object detection actually localize a lesion in a dermoscopic image?
DermiQ is my little experiment with that question.
input → model → “look here”
WHAT I BUILT
DermiQ uses YOLOv8 to detect and localize suspicious lesion regions in dermoscopic images.
Drop an image in. The model runs locally. You get a bounding box around the detected region.
No cloud. No API. Just the experiment running on the machine.
YOLOv8 · Python · Tkinter · PIL · ISIC
THE FUN PART
0.65
I didn't want to just pick a confidence threshold and call it a day. I ran it through multiple experiments.
0.65 felt like the sweet spot.
Low enough
to catch real cases.
High enough
to cut down the noise.
This was the point of the experiment: to test, tune, and learn instead of simply assembling a model.
WHAT SURPRISED ME
The interesting part wasn't getting the model to run.
It was realizing how much the little details mattered: confidence thresholds, dataset inconsistencies, and edge cases.
And how much more useful a location can be than a single yes/no label. The bounding-box output felt more actionable, even while inconsistent labels at the edges made the work messier than it first looked.
WHAT I’M WONDERING NOW
- would RT-DETR localize better?
- what happens with ordinary phone-camera images instead of dermoscopy?
- could the same idea work for other skin conditions?
still haven't answered these.
one more rabbit hole for the collection.
view the code → DermiQ on GitHub — coming soon