How h-ssiqueira/skinlesionai works
Project type and purpose
• This is a research‐oriented machine learning pipeline for skin lesion classification (SkinLesionAI).
• It is organized as Jupyter/Colab notebooks for:
– Experiments (model development, GAN image generation, segmentation)
– Trainings (ConvNeXt, ResNetV2, Xception, Vision Transformer variants)
– Analysis (metrics extraction, graphs, statistical evaluation)
• Data is sourced from the HAM10000 dataset (Kaggle) and results are stored in files and a SQLite database.
File structure
– Top‐level directories:
• Notebooks/experiments – exploratory model development and augmentation
• Notebooks/trainings – systematic training runs, organized by architecture
• analysis/ – metrics extraction, database, visualizations
• docs/ – final PDF report
– Config files: requirements.txt specifies dependencies.
– Database: analysis/database.db and accompanying .sql indicate a metrics storage component.
– No service layers or web frontends – this is a batch‐style pipeline.
README for additional insights
• Objectives: compare models (CNNs and ViT), test techniques (data augmentation, segmentation, GANs, transfer learning)
• Outputs: models (≈22.4 GB), metrics (accuracy, loss, recall, specificity, F1, AUC, confusion matrices)
Main components
a. Data Sources & Storage
• HAM10000 dataset (Kaggle)
• Google Drive (notebooks & model artifacts)
• Local file system & SQLite database (metrics)
b. Notebook Workflows
• Experiment Notebooks (GAN generation, data augmentation, segmentation)
• Training Notebooks (organized by architecture subfolders)
• Analysis Notebooks (read metrics DB, produce graphs)
c. Compute Environments
• Google Colab / TPU/GPU runtimes
• Local Jupyter/Python environment
d. Outputs & Artifacts
• Trained model files
• Serialized weights & checkpoints
• SQLite DB with metrics
• Plots & final PDF report
Relationships and interactions
1. Data ingestion: download from Kaggle → store on Drive/local
2. Experiment phase: notebooks read raw data → generate augmented data (GAN output, segmentations) → save augmented images
3. Training phase: training notebooks read raw+augmented data → train models → write model checkpoints & log metrics to database
4. Analysis phase: analysis notebooks query SQLite DB → generate performance plots → export results to docs
5. Final report compiles graphs and metrics into PDF
Architectural patterns & design principles
• Modular pipeline: separate Experiment, Training, Analysis stages
• File‐based orchestration: Jupyter notebooks drive each stage
• Centralized metrics storage (SQLite) for reproducibility and easy querying
• Use of external compute and storage services (Colab, Drive) for scalability
Diagramming recommendations
– Rectangles for compute modules (notebook clusters)
– Clouds for external services (Kaggle, Google Drive, Colab)
– Raw data → Experiment → Augmented data → Training → Metrics DB → Analysis → Report
Layer 1: External Data & Services
Layer 2: Notebook Stages
Layer 3: Storage & Artifacts
Layer 4: Reporting
Detail without over-splitting
• Seven model families and GAN pipelines under the “Training Notebooks” group, rather than drawing each variant.
• Segmentation and augmentation as sub-components of the Experiment stage.
• Metrics DB is centrally placed to show bidirectional interaction with Training and Analysis.