How Our AI Works
A transparent, step-by-step look at the AI engine behind Tvacha Clinic — purpose-built for Indian dermatology.
50M
Parameters
5 Cr+
Training Images
7×
Per-Photo Passes
2.5×
Cancer Safety Weight
50M parameter ConvNeXt trained on hundreds of thousands of skin images
Advanced Deep Learning Model
50M parameter ConvNeXt architecture — one of the most advanced image classification networks available
Transfer learning from ImageNet-22k for robust feature recognition out of the box
Mixed precision training with gradient accumulation for optimal model convergence
Model Architecture Pipeline
ImageNet-22k
Pre-training
22k classes · rich visual features
Skin Dataset
Fine-tuning
5 Cr+ dermoscopy images
Clinical
Validation
Real-world accuracy testing
50M parameters · ConvNeXt architecture
Mixed precision training · Gradient accumulation
Smart Image Processing
Auto skin detection, quality checks, and clinical-grade normalization
Auto Skin Detection — finds and isolates the skin region using multiple color spaces, works on all skin tones (Fitzpatrick I–VI)
Quality Check — automatically detects blurry, too dark, or overexposed photos and asks for a retake
Clinical-Grade Normalization — bridges the gap between phone photos and clinical images using adaptive contrast enhancement
Background Removal — crops out irrelevant background so the AI focuses only on your skin
Fitzpatrick Scale I–VI Coverage
Processing Stages
Detect
Quality Check
Normalize
Enhance
Trained on real-world phone photo conditions for reliable results
7-Pass Analysis Per Photo
Every photo is analyzed 7 times with different orientations, zoom levels, and lighting adjustments
Results are combined at the mathematical level for higher precision than a single pass
Multi-Photo Mode — upload 3 photos from different angles for even greater accuracy
7-Pass Ensemble Analysis
Combined
Ensemble
Smart Training Pipeline
Focal Loss, balanced sampling, EMA, and early stopping for a robust model
Focal Loss focuses the AI on the hardest-to-distinguish conditions
Balanced class sampling ensures rare conditions are learned equally well
Exponential Moving Average smooths model weights for better generalization
Early stopping prevents overfitting — the AI knows when to stop learning
17 Data Augmentation Strategies
Trained on real-world phone photo conditions for reliable results
Real-world phone photo conditions: compression artifacts, low resolution, partial occlusion, varying lighting and angles
Advanced blending techniques (MixUp & CutMix) create smoother, more generalizable decision boundaries
17 Augmentation Strategies
Evaluation & Transparency
Per-condition precision, recall, and F1 scores — no black box
Evaluated using balanced accuracy across all 13 conditions
Per-condition precision, recall, and F1 scores tracked independently
Dedicated cancer detection rate monitoring
Confidence scores shown for every prediction — no black box
Every prediction includes a confidence score so doctors always know how certain the AI is.
2.5× cancer weighting with triple-layer detection — designed to never miss cancer
Cancer Safety System
2.5× cancer weighting · Triple-layer detection
Cancer classes are weighted 2.5× during training — the AI is trained to never miss cancer
Triple-layer cancer detection: individual class check, grouped probability check, and uncertainty flagging
If combined cancer probability exceeds 15%, you're alerted even if the top result is benign
Low-confidence results return "Uncertain — See a Doctor" rather than a wrong answer
Clinical Questionnaire
5 quick questions that adjust AI predictions using Bayesian statistics
5 quick questions: skin type, age, body location, duration, and symptoms
Adjusts AI predictions using Bayesian statistics based on real clinical data
Example: a changing or growing lesion on sun-exposed skin in a fair-skinned adult increases melanoma weighting
Ready to see it in action?
Start your free trial and explore the AI-powered workflow firsthand.
Start Free TrialNo credit card required