EfficientNet-B0 Architecture

Dermato
Inference.

Classifying 7 types of skin lesions with explainable Grad-CAM heatmaps for early melanoma intervention.

Run Live Demo Back to Portfolio
10,015
HAM10000 Dataset
7
Lesion Classes
224px
Input Resolution
B0
EfficientNet Model
The Medical Imperative

Why AI for Skin Cancer?

Time is Survival

Melanoma has a >99% survival rate if caught early. This drops to <25% once metastasized. AI eliminates the "wait-and-see" delay.

Universal Access

Dermatologists are scarce in underserved regions. Our lightweight B0 model runs on standard laptops, bringing expert-level screening anywhere.

Triage Force Multiplier

By flagging suspicious lesions instantly, AI helps doctors prioritize biopsies for the most critical cases, optimizing clinical workflow.

Skin Lesion Scan
Grad-CAM Active
HAM10000_1001
Melanoma Confidence 88.7%
Biopsy Recommended
Class: Mel
Explainable Healthcare

Transparent
Diagnosis.

Our EfficientNet-B0 backbone isn't just a "black box." With integrated Grad-CAM, we provide visual justifications, showing clinicians exactly which morphological features led to a specific classification.

7-Class Variety

Classifies AKIEC, BCC, BKL, DF, MEL, NV, and VASC with handled imbalanced learning.

Inference Speed

Optimized PyTorch pipeline for real-time mobile and web-based screening via Gradio.

The HAM10000 Matrix

Human Against Machine dataset: 10,015 dermatoscopic images across seven distinct diagnostic categories.

MEL

Melanoma

The most lethal form; early classification is vital.

NV

Melanocytic Nevi

Common moles, dominating 67% of the dataset.

BCC

Basal Cell Carcinoma

Non-melanocytic skin cancer with high cure rates.

AKIEC

Actinic Keratoses

Pre-cancerous lesions appearing on sun-damaged skin.