This paper describes our approach for EAD2019: Multi-class artefact detection in video endoscopy. We optimized focal loss for dense object detection based RetinaNet network pretrained with the ImageNet dataset and applied several data augmentation and hyperparmeter tuning strategies, obtaining a weighted final score of 0.2880 for multi-class artefact detection task and mean average precision (mAP) score of 0.2187 with deviation 0.0770 for multi-class artefact generalisation task. In addition, we developed a U-Net based convolutional neural networks (CNNs) for multi-class artefact region segmentation task and achieved a final score of 0.4320 for the online test set in the competition.
Endoscopic Artefact Detection (EAD) [
For multi-class artefact detection and generalisation tasks, our
solution is based on keras-retinanet [
For multi-class artefact detection task, first of all, we preprocessed the dataset (by resizing the images into 768 1024 pixels), and applied several standard data augmentation techniques including rotation, translation, scaling, and horizontal flipping. We optimized the network with resnet-101 backbone that were pretrained on ImageNet images. Later, we used non-maximum suppression to eliminate some overlapping bounding boxes from predicted bounding boxes as a post-processing step.
In this challenge, the third task was multiclass-artefact generaliation task. Sometimes it is crucial for algorithms to avoid biases induced by specific training dataset. Hence, to be aligned with the organizers’ motivation, we tried to optimize the network that we used for artefact detection task above. Our main intuition was to develop more generalized model so that the model can be used across different endoscopic datasets.
0:4926 0:2880
The second task of the challenge was multi-class artefact
region segmentation. We used an encoder-decoder architecture
called U-Net that is designed for biomedical image
segmentation [
Model performance of multi-class artefact detection task is shown in Table 2. Table 3 shows the overall performance of multi-class artefact generalization task. As explained in Section 2.2, with the limited experiments, our model performance is shown in Table 1 on the final test set of region segmentation task.
In the beginning, when we had phase 1 dataset, we tried to develop our model using 3-fold cross validation. Our models relatively worked as well. Later, when dataset 2 had been released, we incorporated these additional data in our models using 5-fold cross validation. However, our model perform a bit worst. After carefully analyzing, we found that the dataset provided in the second phase is more diverse than the first dataset. We were not able to manage this diversity somehow.
Overall, we noticed a significant gap between our local validation score and the leader board score. Then we reviewed the annotation process more carefully. We found that some cases a bit unusual in the training dataset having more than one classes for almost same bounding boxes. It was understandable why some bounding boxes were overlapped for different class artefacts. However, the situation was not the same for all the bounding boxes of the different/same class(s). An example case is shown in figure 2 (overlapping bounding boxes are marked with circle in yellow color).
The competition was an exciting and educational experience to solve a problem in real-life settings. We thank the organizers for all their hard work for organizing and annotating the datasets for the competition; large medical image data sets of sufficient size and quality for this purpose are rare.
Motivated by the no new-net [