Dataset

Dataset Description

The dataset consists of more than 10,000 images, annotated and verified by medical doctors (experienced endoscopists), including 16 classes showing anatomical landmarks, phatological findings or endoscopic procedures in the GI tract, i.e., with a different number of images in each class. The anatomical landmarks are Z-line, pylorus and cecum, while the pathological finding includes esophagitis, polyps and ulcerative colitis. In addition, we provide two set of images related to removal of polyps, the "dyed and lifted polyp" and the "dyed resection margins". The dataset includes parts of two previously published datasets named Kvasir: A Multi-Class Image Dataset for Computer Aided Gastrointestinal Disease Detection and Nerthus: A Bowel Preparation Quality Video Dataset. In addition to this we will also provide several videos for the task of creating automatic reports. Videos will contain the same classes as the image dataset.

Classes used in the dataset are:

• blurry-nothing
• colon-clear
• dyed-lifted-polyps
• dyed-resection-margins
• esophagitis
• instruments
• normal-cecum
• normal-pylorus
• normal-z-line
• out-of-patient
• polyps
• retroflex-rectum
• retroflex-stomach
• stool-inclusions
• stool-plenty
• ulcerative-colitis

The images of the dataset come from different equipment installed in different Norwegian hospitals with different resolutions from 720x576 up to 1920x1072 pixels and is organized in a way where they are sorted in separate folders named accordingly to the content.

Some of the included classes of images have a green sub-picture in the image illustrating the position and configuration of the endoscope inside the colon, by use of an electromagnetic imaging system (ScopeGuide, Olympus Europe) that may support the interpretation of the image.

The whole dataset is split onto two equally sized development and test datasets. Pre-extracted features for all data (visual) will also be provided. The ground truth for the data is collected from the medical experts annotations.

Dataset Details

Both the development and the test datasets consists of images sorted to classes with different number of images per each class stored in two archives: images archive and features archive.

The images archive contains raw images sorted to classes with different number of images per each class. In the development dataset the images are stored in the separate folders named accordingly to the name of the class images belongs to. In the test dataset all the images stored in one folder. The image files are encoded using JPEG compression. The encoding settings can vary across the dataset and they reflecting the a priori unknown endoscopic equipment settings. The extension of the image files is ".jpg".

The features archive contains the extracted visual feature descriptors for all the image from the images archive. The extracted visual features are stored in the separate folders and files named accordingly to the name and the path of the corresponding image files. The extracted visual features are the global image features, namely: JCD, Tamura, ColorLayout, EdgeHistogram, AutoColorCorrelogram and PHOG. Each feature vector consists of a number of floating point values. The size of the vector depends on the feature. The size of the feature vectors are: 168 (JCD), 18 (Tamura), 33 (ColorLayout), 80 (EdgeHistogram), 256 (AutoColorCorrelogram) and 630 (PHOG). The extracted visual features are stored in the text files. Each file consists of eight lines, one line per each feature. Each line consists of a feature name separated from the feature vector by colon. Each feature vector consists of a corresponding number of floating point values separated by commas. The extension of the extracted visual feature files is ".features".