=Paper=
{{Paper
|id=Vol-2366/EAD2019_paper_9
|storemode=property
|title=Multi-class Artefact Detection In Video Endoscopy Via Convolution
Neural Networks
|pdfUrl=https://ceur-ws.org/Vol-2366/EAD2019_paper_9.pdf
|volume=Vol-2366
|authors=Mohammad Azam Khan, Jaegul Choo
}}
==Multi-class Artefact Detection In Video Endoscopy Via Convolution
Neural Networks ==
MULTI-CLASS ARTEFACT DETECTION IN VIDEO ENDOSCOPY VIA CONVOLUTION NEURAL NETWORKS Mohammad Azam Khan, Jaegul Choo Korea University Department of Computer Science and Engineering Seoul, South Korea. {a khanss, jchoo}@korea.ac.kr ABSTRACT 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 pre- trained 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 detec- tion 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 seg- mentation task and achieved a final score of 0.4320 for the Fig. 1. Overall detection pipeline for multi-class artefact de- online test set in the competition. tection and generalisation task. Index Terms— Endoscopic artefact, Video endoscopy, artefact generalization, Convolutional neural networks was really appealing to us for its training simplicity. Over- all detection pipeline for two tasks is shown in Figure 1. 1. INTRODUCTION Endoscopic Artefact Detection (EAD) [1, 2] is a core chal- lenge in facilitating diagnosis and treatment of diseases in 2.1. Multi-class artefact detection and generalisation hollow organs. This Challenge highlights the growing appli- cation of artificial intelligence (AI) in general, and specific For multi-class artefact detection task, first of all, we prepro- application of deep learning (DL) techniques for the early de- cessed the dataset (by resizing the images into 768 × 1024 tection of numerous cancers, therapeutic procedures and min- pixels), and applied several standard data augmentation tech- imally invasive surgery. In this concern, the organizers mainly niques including rotation, translation, scaling, and horizontal focused on three sub-tasks for this challenge using the EAD flipping. We optimized the network with resnet-101 back- dataset [1, 2]: multi-class artefact detection, region segmen- bone that were pretrained on ImageNet images. Later, we tation and detection generalization. used non-maximum suppression to eliminate some overlap- ping bounding boxes from predicted bounding boxes as a 2. OUR APPROACH post-processing step. In this challenge, the third task was multiclass-artefact For multi-class artefact detection and generalisation tasks, our generaliation task. Sometimes it is crucial for algorithms to solution is based on keras-retinanet [3] which is basically an avoid biases induced by specific training dataset. Hence, to implementation of a popular dense object detection method be aligned with the organizers’ motivation, we tried to op- called RetinaNet [4] using open-source framework Keras [5] timize the network that we used for artefact detection task with Tensorflow1 back-end. The RetinaNet is a single-stage above. Our main intuition was to develop more generalized convolutional neural network detection architecture, which model so that the model can be used across different endo- 1 https://www.tensorflow.org/ scopic datasets. Table 1. Segmentation scores. Dataset Model Overlap F2 Final Test (Online) U-Net 0.4324 0.4310 0.4320 Table 2. Detection results. Dataset Backbone mAP IoU Score Validation ResNet101 0.4547 0.5167 0.4926 Online ResNet101 0.2581 0.3330 0.2880 Table 3. Generalisation results. Dataset Backbone mAP Dev Test (online) ResNet101 0.2187 0.0770 Fig. 2. Sample image having almost same bounding boxes for different classes. 2.2. Multi-class artefact region segmentation The second task of the challenge was multi-class artefact re- than one classes for almost same bounding boxes. It was un- gion segmentation. We used an encoder-decoder architecture derstandable why some bounding boxes were overlapped for called U-Net that is designed for biomedical image segmenta- different class artefacts. However, the situation was not the tion [6]. The encoder path identifies the contents of the image same for all the bounding boxes of the different/same class(s). while the decoder part localize where the contents are avail- An example case is shown in figure 2 (overlapping bounding able. More importantly, in a U-Net, the output is an image boxes are marked with circle in yellow color). with the same dimension of the input, but with one channel. The competition was an exciting and educational experi- Unfortunately, we were not able to make extensive experi- ence to solve a problem in real-life settings. We thank the ments for this task. organizers for all their hard work for organizing and annotat- ing the datasets for the competition; large medical image data 3. RESULTS sets of sufficient size and quality for this purpose are rare. Model performance of multi-class artefact detection task is shown in Table 2. Table 3 shows the overall performance of 5. CONCLUSION multi-class artefact generalization task. As explained in Sec- tion 2.2, with the limited experiments, our model performance Motivated by the no new-net [7], we wanted to demonstrate is shown in Table 1 on the final test set of region segmentation the effectiveness of well trained state-of-the-art networks in task. the context of three different tasks of EAD 2019 challenge. While most of the researchers are currently besting each other with minor modification of exiting networks, we instead fo- 4. DISCUSSION cused on the training process. The detection of specific arte- facts and then precise boundary delineation of detected arte- In the beginning, when we had phase 1 dataset, we tried to facts, and finally detection generalization of independent of develop our model using 3-fold cross validation. Our models specific data type and source - all would mark critical steps relatively worked as well. Later, when dataset 2 had been forward for this domain. 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 6. REFERENCES provided in the second phase is more diverse than the first dataset. We were not able to manage this diversity somehow. [1] Sharib Ali, Felix Zhou, Christian Daul, Barbara Braden, Overall, we noticed a significant gap between our local Adam Bailey, Stefano Realdon, James East, Georges validation score and the leader board score. Then we re- Wagnires, Victor Loschenov, Enrico Grisan, Walter Blon- viewed the annotation process more carefully. We found that del, and Jens Rittscher, “Endoscopy artifact detection some cases a bit unusual in the training dataset having more (EAD 2019) challenge dataset,” 2019. [2] Sharib Ali, Felix Zhou, Adam Bailey, Barbara Braden, James East, Xin Lu, and Jens Rittscher, “A deep learning framework for quality assessment and restoration in video endoscopy,” CoRR, vol. abs/1904.07073, 2019. [3] Hans Gaiser et al., “Fizyr/keras-retinanet: 0.5.0,” 2018. [4] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar, “Focal loss for dense object detection,” in 2017 IEEE International Conference on Computer Vi- sion (ICCV). oct 2017, IEEE. [5] François Chollet et al., “Keras,” https://github.com/fchollet/keras, 2015, GitHub. [6] Olaf Ronneberger, Philipp Fischer, and Thomas Brox, “U-net: Convolutional networks for biomedical image segmentation,” in Lecture Notes in Computer Science, pp. 234–241. Springer International Publishing, 2015. [7] Fabian Isensee, Philipp Kickingereder, Wolfgang Wick, Martin Bendszus, and Klaus H. Maier-Hein, “No new- net,” in Brainlesion: Glioma, Multiple Sclerosis, Stroke and Traumatic Brain Injuries, pp. 234–244. Springer In- ternational Publishing, 2019.