Overview

This pipeline is modified from the 4th place solution of RANZCR CLiP - Catheter and Line Position Challenge in Kaggle: https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification

The original solution is produced by Team Watercooled, and the authors are Dieter (https://www.kaggle.com/christofhenkel) and Psi (https://www.kaggle.com/philippsinger).

Preparations

Recommend to use MONAI docker:

docker pull projectmonai/monai:latest

Otherwise, except for monai required libraries, please ensure that opencv-python and scipy are installed. The command could be:

pip install opencv-python
pip install scipy

Please download the dataset from the competition site and modify the dataset path in configs/default_config.py.

In this pipeline, the data split file train_folds.csv is achieved from the following public kernel which is different from the original solution: https://www.kaggle.com/underwearfitting/how-to-properly-split-folds Please download and put it into the dataset's path

GPU used:

A single NVIDIA Tesla V100 32G

Inference

To do inference locally, the command could be:

python train.py -c cfg_seg_40_1024d_full -i True -p your_actual_local_weight_folder_path

In Kaggle, since the kernel for submission does not support the internet access, you may need to copy the code of MONAI/monai/ (if using MONAI docker, the path is /opt/monai/monai/) into this directory and upload the whole folder as a Kaggle dataset to use. Please take the run_infer function in train.py for reference and write the script.

Places uncovered

The original solution is consisted with 6 different models/training strategies:

1. Backbone: EfficientNet-b8, with imagenet pretrained weights
2. Backbone: EfficientNet-b8, with AdvProp (https://rwightman.github.io/pytorch-image-models/models/advprop/) pretrained weights
3. Backbone: EfficientNet-b7, with imagenet pretrained weights
4. Backbone: EfficientNet-b7, with AdvProp pretrained weights
5. Backbone: EfficientNet-b7, with noisy student(https://rwightman.github.io/pytorch-image-models/models/noisy-student/) pretrained weights
6. Backbone: EfficientNet-b7, with noisy student pretrained weights (different training strategies)

However, only the 2nd, 3rd and 4th methods are covered in the reproduction. The reason is that the EfficientNet implemented in MONAI refered to EfficientNet-PyTorch (https://github.com/lukemelas/EfficientNet-PyTorch) rather than pytorch-image-models (https://github.com/rwightman/pytorch-image-models Team Watercooled refers to this repository). EfficientNet-PyTorch does not provide the uncovered pretrained weights, and due to some differences in efficientnet structure, we cannot make use of weights from pytorch-image-models instead.

Performance

According to the above mentioned reasons, this pipeline does not fully reproduce the original solution. As for the models that have the suitable pretrained weights, comparable performances are reachable.

Training with a single efficientnet-b8 model with AdvProp pretrained weights can achieve:

1. one seed can reach to around 0.97013 for private LB and 0.97092 for public LB.
2. four seeds can reach to around 0.97429 for private LB and 0.97460 for public LB.

The commands are:

python train.py -c cfg_seg_40_1024d_full -s 657028
python train.py -c cfg_seg_40_1024d_full -s 770799
python train.py -c cfg_seg_40_1024d_full -s 825460
python train.py -c cfg_seg_40_1024d_full -s 962001

In addition, ensemble with more models can get a higher score:

python train.py -c cfg_seg_philipp_16_ch_1024_ap_full -s 868472
python train.py -c cfg_seg_philipp_16_ch_1024_ap_full -s 183105

python train.py -c cfg_seg_philipp_16_ch_1024_nons_full -s 701922
python train.py -c cfg_seg_philipp_16_ch_1024_nons_full -s 7259