edziocodes/medgemma-breast-cancer
The edziocodes/medgemma-breast-cancer model is a 4.3 billion parameter vision-language model, fine-tuned from Google's MedGemma-4B-IT, specifically for binary breast cancer classification from 2D mammogram images. It was adapted using LoRA on a balanced subset of the OMAMA 256x256 dataset, achieving 98.88% accuracy and 99.66% sensitivity in identifying cancer. This model is optimized for medical imaging analysis, particularly for reducing false negatives in breast cancer screening applications.
Loading preview...
Overview
This model, edziocodes/medgemma-breast-cancer, is a specialized fine-tuned version of Google's MedGemma-4B-IT, designed for binary classification of breast cancer from 2D mammogram images. It leverages a 4.3 billion parameter architecture and was fine-tuned using LoRA on a balanced subset of the OMAMA 256x256 dataset, ensuring equal representation of cancer and non-cancer cases.
Key Capabilities
- High Accuracy: Achieves 98.88% accuracy on a balanced test set.
- Exceptional Sensitivity: Demonstrates 99.66% sensitivity, significantly reducing false negatives (from 243 to 5 cases compared to baseline).
- Medical Image Analysis: Specifically adapted for
image-text-to-texttasks in mammography. - Robust Performance: Outperforms the MedGemma-4B-IT baseline by 7.28 percentage points in accuracy and 16.18 percentage points in sensitivity for this specific task.
Good For
- Research and Education: Ideal for studies in medical imaging and breast cancer classification.
- Screening Applications: Its high sensitivity makes it valuable for identifying potential cancer cases, where missing a diagnosis has severe consequences.
- Further Development: Provides a strong foundation for building more advanced diagnostic tools, with the understanding that it is not a certified clinical diagnostic tool on its own.