My UNet++ implementation : pneumothorax classification & segmentation
On the left: ground truth images & labels. No PT present means there's no pneumothorax. If there is a ground truth pneumothorax, the radiologist's label has been coloured in red. On the right: my model's predictions. Predicted pnuemothorax is highlighted, with yellow being the highest confidence areas. Pause the video if you want to take a longer look at a pair of images. I wrote UNet (https://arxiv.org/pdf/1505.04597) and UNet++ (https://arxiv.org/pdf/1912.05074) from scratch in Tensorflow 2.3. Then, I trained a UNet++ model on the 2018 SIIM-ACR Kaggle pneumothorax data (https://www.kaggle.com/c/siim-acr-pneumothorax-segmentation) and got a pretty good result! This video shows 50 examples. The codebase I wrote is for a segmentation-classification pipeline and it has lots of features such as data augmentation, precision-recall and dice score calculations etc. Check it out here: https://github.com/albertsokol/pneumothorax-detection-unet
On the left: ground truth images & labels. No PT present means there's no pneumothorax. If there is a ground truth pneumothorax, the radiologist's label has been coloured in red. On the right: my model's predictions. Predicted pnuemothorax is highlighted, with yellow being the highest confidence areas. Pause the video if you want to take a longer look at a pair of images. I wrote UNet (https://arxiv.org/pdf/1505.04597) and UNet++ (https://arxiv.org/pdf/1912.05074) from scratch in Tensorflow 2.3. Then, I trained a UNet++ model on the 2018 SIIM-ACR Kaggle pneumothorax data (https://www.kaggle.com/c/siim-acr-pneumothorax-segmentation) and got a pretty good result! This video shows 50 examples. The codebase I wrote is for a segmentation-classification pipeline and it has lots of features such as data augmentation, precision-recall and dice score calculations etc. Check it out here: https://github.com/albertsokol/pneumothorax-detection-unet