Convolutional Neural Networks for Pneumothorax Segmentation

This repository contains code for my approach to the 2019 SIIM-ACR Pneumothorax Segmentation Kaggle Challenge. Implemented using tensorflow.keras.

Credits to Amazingly Helpful Kernels

• Preprocessed dataset from this kernel by lafoss
• Unet Plus Plus with EfficientNet Encoder and Unet with EfficientNet Encoder in Keras by Siddhartha

Preprocessing

The only preprocessing that was done was CLAHE and rescaling the intensities to [0, 255]. The output images were converted to .png files.

Solution

[0.8523 Public LB // 0.8980 Private (1% of test data)]
Note: Models were not retrained on the stage 1 test labels.
My final solution was essentially just running the Unet Plus Plus with EfficientNet Encoder kernel 10 different times and ensembling the SWA snapshots with horizontal flipping for test-time augmentation and zeroing out predictions with small ROIs (pneumothorax_seg.inference.segmentation_only.py). There were some variations in the models I ensembled (mainly, the data augmentation), but they didn't have much impact on the overall performance across folds. Here are the model weights I used.

What about the other models?

Classification-Segmentation Cascade

[Best: 0.8500 Public LB]
With this approach, I was planning on replicating something similar to the 1st Place Solution to the RSNA Pneumonia Detection Challenge, where they used ensembles of classification models (5 10-fold CV ensembles) in conjuction with ensembles of detection models. Both that challenge and this one used the NIH ChestXray14 Dataset, and they also released pre-trained weights on said dataset.

• used binary cross entropy, Adam, SWA, and a cosine annealing LR.
• Data augmentation was horizontal flipping, color inversion, gaussian smoothing, rotations, zooms, and random gamma (as shown here)

I found that I couldn't really get any real worthwhile performance improvements with a small ensemble of 4 classification models on a single fold (EfficientNetB4 with ImageNet weights; DenseNet169, InceptionResNetV2, and Xception on NIH weights) and the UEfficientNetB4 trained on pneumothorax positive patients only, so I just reverted back to segmentation only approaches. (Classification F-Scores were around 0.7-0.76 and AUCs around 0.82-0.87 depending on the thresholds). It would probably do better with larger ensembles, but time was a serious constraint with limited computing power.

However, I did create a cleaner version found in the original repository to load the NIH weights in the classification models, which is found in pneumothorax_seg.training.utils.load_pretrained_classification_model.

Misc. Approaches

• Regular U-Net: Just did a basic recursive tf.keras implementation with conv, LeakyReLU, and Instance Normalization. However, it didn't do too well (CV < 0.75), so I opted to just use the U-Net++. (This was done on 512x512 inputs).
• VAE CNN: This architecture is from an unofficial implementation of the 1st place solution to the 2018 MICCAI Brain Tumour Segmentation Challenge. I just made a 2D version so that it fit the code style in this repository and that it worked for tensorflow.keras and channels_last. I didn't play with it much, but it did about as well as the vanilla U-Net, while consuming a lot more compute time. (This was also done on 512x512 inputs). Better hyperparameter tuning might've made a bigger difference, but I didn't have much compute to work with.