yrribeiro / detechest-ood

testing vlm for detecting out-of-distribution images in chest x-ray datasets

Work in Progress

VLMs for OOD Chest X-ray images detection.

Mentor must see

• Achieve 96% f1-score with CheXZero Image Embeddings + Isolation Forest. check the ipynb
• Improved AUC scores for each class (51% normal, 66% exclude) in zero-shot classification by detailing the model's input prompt. check the ipynb

Data

All data was acquired from the Padchest dataset. 1932 images splitted in 1821 'normal' images and 111 'exclude' images, a.k.a, ood images.