Strong pixel level overlap on the held out project test set.
PROJECT 09 / Computer Vision · Medical Imaging
Skin Lesion Segmentation
A medical imaging prototype that traces the pixel level boundary of a skin lesion, helping show its shape and extent rather than returning only a category name.
DATASETHAM10000 Lesion Segmentations
SIZE10,015 image and mask pairs
MEASURED EVIDENCE
What can be verified.
Metrics and outputs drawn from the project artifacts—not estimates added for presentation.
Best recorded overlap during 20 training epochs.
A U-Net trained from scratch without a pretrained backbone.
Split into 8,012 train, 1,001 validation, and 1,002 test pairs.
VALIDATION
The public repository documents the architecture, data split, loss function, training configuration, segmentation visualizations, and live Hugging Face demonstration.
Why this project matters
For some image tasks, knowing that a lesion is present is not enough; researchers may also need to know exactly where it begins and ends. This project focused on turning a dermoscopic image into a detailed lesion mask.
What was developed
The implementation trained a 31 million parameter U-Net from scratch for 20 epochs using 10,015 HAM10000 image and mask pairs. The application produces a predicted mask and an overlay so users can visually compare the outlined region with the original image.
What it means
The model achieved a 0.9115 test Dice score, indicating strong overlap on the project test data. The result demonstrates how segmentation can measure an object's precise area, while still requiring broader clinical validation before real world use.
04 / WHERE THIS WORK APPLIES
From project to practical use.
Segmentation is useful in medical image measurement, treatment monitoring, surgical planning research, satellite and land use mapping, crop analysis, defect inspection, autonomous systems, and any task requiring precise object boundaries.
05 / LIMITATION
What the project does not solve yet.
Training on a single public dataset may limit generalization across devices, skin tones, and clinical settings, especially around ambiguous lesion boundaries.
06 / NEXT ITERATION
Future advancement
Use multi center datasets, stronger augmentation and ensemble methods, uncertainty maps, and clinician reviewed external validation.
TECHNICAL TOOLKIT
U-NetHAM10000PyTorchImage SegmentationGradio