Multi-Label Chest X-ray Classifier
Budget: $8 – $15 USD
I have a very large collection of JPEG chest X-ray images, each already tagged with up to 14 possible findings. I want a PyTorch-based model that combines a proven CNN backbone with a Vision Transformer head (or hybrid) to deliver reliable multi-label classification.
Scope
• Design and implement the architecture in PyTorch, leveraging transfer learning where sensible.
• Handle class imbalance and dataset split logic (train/val/test) for this multi-label scenario.
• Train the network end-to-end, monitor AUC, F1 and other relevant metrics, and iterate until performance is consistently strong.
• Provide clean, well-commented code plus a short README so I can reproduce results on my own machine.
• Supply final trained weights and an inference script that accepts a folder of JPEGs and outputs the 14-class probability vector for each image.
• Summarise key training choices, hyper-parameters and results in a concise report (a page or two is enough).
The dataset is already labeled, so you can go straight to exploration and modelling. If you have experience mixing CNN feature extractors with transformers for medical imaging, I’d love to see it.
Scope
• Design and implement the architecture in PyTorch, leveraging transfer learning where sensible.
• Handle class imbalance and dataset split logic (train/val/test) for this multi-label scenario.
• Train the network end-to-end, monitor AUC, F1 and other relevant metrics, and iterate until performance is consistently strong.
• Provide clean, well-commented code plus a short README so I can reproduce results on my own machine.
• Supply final trained weights and an inference script that accepts a folder of JPEGs and outputs the 14-class probability vector for each image.
• Summarise key training choices, hyper-parameters and results in a concise report (a page or two is enough).
The dataset is already labeled, so you can go straight to exploration and modelling. If you have experience mixing CNN feature extractors with transformers for medical imaging, I’d love to see it.