Measured across seven HAM10000 lesion classes.
PROJECT 07 / Computer Vision · Explainability
Skin Disease Classifier
An explainable image classification prototype that suggests a skin lesion category and highlights the image regions that most influenced the prediction.
DATASETHAM10000
SIZE10,015 dermoscopic images · 7 classes
MEASURED EVIDENCE
What can be verified.
Metrics and outputs drawn from the project artifacts—not estimates added for presentation.
Strongest documented class level result.
Reported class level detection result on the project test split.
From 10,015 total dermoscopic images.
VALIDATION
The repository includes the test split, class level results, training curves, confusion matrix, and Grad-CAM explainability outputs. It remains a research prototype, not a diagnostic system.
Why this project matters
An image model can produce a label without showing why, which makes the result difficult to examine or trust. This project arose from the need to pair a skin image prediction with a visual clue about where the model focused.
What was developed
The implementation adapted EfficientNetB0 through transfer learning on the HAM10000 dataset and added Grad-CAM heatmaps. The interface returns both a predicted class and an overlay that shows which parts of the image contributed most strongly.
What it means
The prototype makes computer vision behavior easier to inspect and discuss. The heatmap is not a medical explanation or diagnosis, but it helps reveal when a model may be attending to a lesion, or to irrelevant image artifacts.
04 / WHERE THIS WORK APPLIES
From project to practical use.
Similar explainable classification methods can assist research in dermatology, radiology, pathology, manufacturing quality control, crop disease identification, insurance inspection, and other image review workflows.
05 / LIMITATION
What the project does not solve yet.
Predictions can be sensitive to image quality and dataset representation, and Grad-CAM indicates attention rather than proving clinical reasoning.
06 / NEXT ITERATION
Future advancement
Train on more diverse clinical imagery, improve calibration and uncertainty reporting, and validate performance with dermatology experts before clinical use.
TECHNICAL TOOLKIT
EfficientNetB0Transfer LearningGrad-CAMGradioTensorFlow