CheXNet Multi-Disease Accuracy Upgrade

I need an experienced deep-learning practitioner to push CheXNet beyond its current “enhanced” release so it can reliably distinguish five classes on chest X-rays: Cardiomegaly, COVID-19, Normal, Pneumonia and Tuberculosis. I will supply a mixed dataset that combines well-known open-source repositories with my own curated studies. Before training, please run the full pre-processing chain—normalization, smart data augmentation and consistent resizing—so that class balance and resolution do not become hidden biases. The baseline to beat is the latest public enhanced-CheXNet checkpoint. I care most about overall accuracy, so report that headline figure on a held-out test split, but also include supporting curves and a confusion matrix so we can verify class-level behaviour. Deliverables • Updated model weights and reproducible training script (PyTorch preferred, Keras acceptable) • Inference notebook or API stub that takes a single DICOM or PNG and returns the predicted label with probability scores • Brief technical note comparing your model to the baseline, highlighting architecture tweaks, hyper-parameters and accuracy gains Acceptance is contingent on demonstrably higher accuracy than the enhanced CheXNet benchmark when both are evaluated on the same unseen test set I will provide.

Project ID: 39719586