Dimaodessa/pneumonia-cxr

This model performs binary classification and segmentation for pneumonia (lung opacity) in frontal chest radiographs. It is a tf_efficientnetv2_s backbone with a U-Net decoder and linear classification head. The model was trained on the RSNA Pneumonia Detection Challenge dataset and the SIIM-FISABIO-RSNA COVID-19 Detection dataset. Both of these datasets were annotated with bounding boxes, which were converted to ellipsoid segmentation masks.

Classification performance on a holdout test set of 1,334 images from the RSNA dataset and 317 images from the SIIM-FISABIO-RSNA dataset:

RSNA + SIIM-FISABIO-RSNA (n=1,651): AUC 0.900
                    RSNA (n=1,334): AUC 0.885
       SIIM-FISABIO-RSNA (n=317)  : AUC 0.914