The human digestive system is prone to suffer from many different diseases and abnormalities throughout a human lifetime. Some of these may be life-threatening and pose a serious risk to a patient's health and well-being. In most cases, if the detection of lethal disease is done early enough, it can be treated with a high chance of being fully healed. Therefore, it is important that all lesions are identified and reported during a routine investigation of the GI tract. Currently, the gold-standard in performing these investigations is through video endoscopies, which is a procedure involving a small camera attached to a tube that is inserted either orally or rectally. However, there is one major downside to this procedure. The method is highly dependent on the skill and experience of the person operating the endoscope, which in turn results in a high operator variation and performance. This is one of the reasons for high miss-rates when measuring polyp detection performance, with some miss-rates being as high as 20%. We see this as an opportunity to aid medical doctors by helping them detect lesions through automatic frame analysis done live during endoscopy examinations. The pattern recognition community has a lot of knowledge which could assist in this task, making it a perfect fit for ICPR. The work done in this competition has the potential of making a real societal impact, as it directly affects the quality of care that health-care professionals can provide.
The challenge consists of three tasks, each targeting a different requirement for in-clinic use. The first task involves classifying images from the GI tract into 23 distinct classes. The second task focuses on efficiant classification measured by the amount of time spent processing each image. The last task relates to automatcially segmenting polyps. More information can be found on the task page.
The dataset can be split into four distinct parts; Labeled image data, unlabeled image data, segmented image data, and annotated video data. Each part is further described below. In total, the dataset contains 110,079 images and 373 videos where it captures anatomical landmarks and pathological and normal findings. The results is more than 1.1 million images and video frames all together.Labeled Images Unlabeled Images Segmented Images Videos
The dataset is split into two distinct parts; the classification dataset and segmentation dataset. The classification dataset should be used to perform the detection and speed tasks, while the segmentation part should be used for the segmentation task.Classification Dataset Segmentation Dataset
Labeled Images. In total, the dataset contains 10,662 labeled images stored using the JPEG format. The images can be found in the images folder. The classes, which each of the images belongto, correspond to the folder they are stored in (e.g., the "polyp" folder contains all polyp images, the "barretts" folder contains all images of Barrett’s esophagus, etc.). The number of images per class are not balanced, which is a general challenge in the medical field due to the fact that some findings occur more often than others. This adds an additional challenge for researchers, since methods applied to the data should also be able to learn from a small amount of training data. The labeled images represent 23 different classes of findings.
Unlabeled Images. In total, the dataset contains 99,417 unlabeled images. The unlabeled images can be found in the unlabeled folder which is a subfolder in the image folder, together with the other labeled image folders. In addition to the unlabeled image files, we also provide the extracted global features and cluster assignments in the Hyper-Kvasir Github repository as Attribute-Relation File Format (ARFF) files. ARFF files can be opened and processed using, for example, the WEKA machine learning library, or they can easily be converted into comma-separated values (CSV) files.
Annotated Videos. The dataset contains a total of 373 videos containing different findings and landmarks. This corresponds to approximately 11.62 hours of videos and 1,059,519 video frames that can be converted to images if needed. Each video has been manually assessed by a medical professional working in the field of gastroenterology and resulted in a total of 171 annotated findings.