Melanoma Detection

Image classifiers are difficult to assess outside a notebook. This project turns the trained model into an interactive workflow where a user can upload a lesion image, preview it, and request a benign-or-malignant classification.

The work uses lesion images from the International Skin Imaging Collaboration 2020 challenge. It is a machine-learning research demonstration, not a clinical diagnostic tool.

The Streamlit application keeps the inference path explicit and reproducible from upload to result.

1. Start with transfer learning. A pretrained EfficientNetB6 extracts visual features, while a bidirectional LSTM forms the hybrid classification model documented in the project.
2. Make deployment portable. The trained Keras model is converted into a single hybrid.tflite artifact consumed directly by the application.
3. Keep preprocessing consistent. Every uploaded image follows the same resize, array conversion, pixel normalization, and batch-expansion sequence used by inference.
4. Expose the model through a narrow interface. Streamlit handles upload, preview, and the explicit classify action without requiring a separate frontend or API.