{"paper_id":"c0bbe79c-d2d5-43f8-aa12-cd6c9c41018f","body_text":"GLENDA: Gynecologic Laparoscopy\nEndometriosis Dataset\nAndreas Leibetseder1[0000−0002−9535−966X], Sabrina Kletz 1[0000−0002−7275−0594],\nKlaus Schoeffmann1[0000−0002−9218−1704], Simon Keckstein2, and\nJ¨ org Keckstein3\n1 Institute of Information Technology, Klagenfurt University,\nKlagenfurt 9020, Austria\n{aleibets,sabrina,ks}@itec.aau.at\n2 University Hospital, Ludwig-Maximilians-University Munich,\nMunich 80799, Germany\nsimon.keckstein@med.uni-muenchen.de\n3 Medical Faculty, Ulm University,\nUlm 89081, Germany\njoerg@keckstein.at\nAbstract. Gynecologic laparoscopy as a type of minimally invasive sur-\ngery (MIS) is performed via a live feed of a patient’s abdomen survey-\ning the insertion and handling of various instruments for conducting\ntreatment. Adopting this kind of surgical intervention not only facili-\ntates a great variety of treatments, the possibility of recording said video\nstreams is as well essential for numerous post-surgical activities, such\nas treatment planning, case documentation and education. Nonetheless,\nthe process of manually analyzing surgical recordings, as it is carried out\nin current practice, usually proves tediously time-consuming. In order\nto improve upon this situation, more sophisticated computer vision as\nwell as machine learning approaches are actively developed. Since most\nof such approaches heavily rely on sample data, which especially in the\nmedical field is only sparsely available, with this work we publish the\nGynecologic Laparoscopy ENdometriosis DAtaset (GLENDA) – an im-\nage dataset containing region-based annotations of a common medical\ncondition named endometriosis, i.e. the dislocation of uterine-like tissue.\nThe dataset is the first of its kind and it has been created in collaboration\nwith leading medical experts in the field.\nKeywords: lesion detection · endometriosis localization · medical dataset\n· region-based annotations · gynecologic laparoscopy\n1 Introduction\nMinimally invasive surgery (MIS) considerably reduces trauma inflicted upon\npatients during medical interventions, since, as opposed to traditional open sur-\ngery, treatments are applied less intrusively. As a typical form of MIS,endoscopy\nis performed by inserting a small camera, the endoscope, as well as a variety of\narXiv:2508.21398v1  [cs.CV]  29 Aug 2025\n\n2 A. Leibetseder et al.\ninstruments into the human body via natural or artificially created orifices. In\nthe case of gynecologic laparoscopy such incisions are created into the abdomen\nin order to treat conditions related to the female reproductive system. The ac-\ncordingly obtained video feed of an individual’s inner anatomy is projected onto\nexternal monitors providing physicians with adequate visuals for performing sur-\ngery.\nWith the prospect of conducting surgeries in such a manner comes the pos-\nsibility of recording entire procedures, an opportunity that is in fact pursued by\nmost modern medical facilities. Apart from representing valuable evidence for\nlawful investigations, these kind of recordings more importantly are consulted\nby medical practitioners for further treatment planning, case revisitations or\neven educational purposes. Seemingly a convenient improvement, several down-\nsides, however, considerably diminish the usefulness of archived surgery footage:\nrecording the typically hours-long surgeries filmed in high-definition on a daily\nbasis requires elaborate long-term storage solutions. Furthermore, in order to\nremain useful, video archives of such magnitudes must easily be searchable even\nby potentially non tech-savvy staff. The consequentially arising need for more\nsophisticated systems capable of aiding physicians post- as well as even intra-\nsurgery creates great opportunities and challenges for various scientific commu-\nnities, not least the ones concerned with multimedia [7].\nAlthough machine learning has successfully been applied in the field of med-\nical imaging [5], specifically for the task of disease classification and diagnosis,\nmuch needed published datasets for feeding corresponding algorithms are only\nsparsely available. This not only is due to the increased sensitiveness of such data\nbut as well a consequence of the broad spectrum of different imaging technolo-\ngies4 utilized for a great variety of purposes. When merely regarding endoscopy\nin general the number of publically available datasets is reduced even more dras-\ntically, leaving only a few for the sub-discipline of laparoscopy: Cholec80 [11],\nLapChole [9], GI dataset [12], SurgicalActions160 [8] and LapGyn4 [4] to name\nsome recent ones.\n(a) rASRM examples (peritoneum, ovary)\nof varying severity\n (b) Enzian examples\nFig. 1: Example endometriosis locations for rASRM (a) and Enzian (b).\nAs cholecystectomy, i.e. the removal of the gallbladder, is the most fre-\nquently conducted laparoscopic surgery [10], released datasets commonly are\ncreated from this procedure. Keeping that in mind, with this work we specif-\n4 e.g. X-ray, computed tomography (CT) scans, magnetic resonance imaging (MRI),\nultrasound, ...\n\nGLENDA: Gynecologic Laparoscopy Endometriosis Dataset 3\nically target a different kind of procedure, which typically as well is treated\nlaparoscopically: the diagnosis, inspection and surgical removal of endometriosis\n– a benign but painful anomaly among women in child bearing age involving\nthe growth of uterine-like tissue in locations outside of the uterus. The condi-\ntion can be found in various positions and severities, often in multiple instances\nper patient requiring a physician to determine its extent. This most frequently\nis accomplished by calculating its magnitude via utilizing the combination of\ntwo popular classification systems, the revised American Society for Reproduc-\ntive Medicine (rASRM) score [2] and the European [1] Enzian classification [3],\nwhich describe the anomaly’s potential anatomical location and severity on a\nthree-level scale. Figure 1 shows a few examples of these locations as described\nby both of these systems. Our contribution, the Gynecologic Laparoscopy EN-\ndometriosis DAtaset5 (GLENDA) dataset comprises a subset of these locations\nand has been created with leading medical experts in the field of endometriosis\ntreatment. The dataset and with it our contribution can be characterized as\nfollows:\nSource 300+ video segments and frames selected from a pool of 400+ individual\nfull surgery videos.\nImages 25K+ Images, consisting of 12K+ positive, i.e. pathological images as-\nsociated with endometriosis, and 13K+ negative examples, i.e. non-pathological\nimages without visible endometriosis.\nAnnotations 500+ hand-drawn region-based class-specific endometriosis an-\nnotations on 300+ images/keyframes.\nClasses Five pathological categories, of which four are based on the location of\nthe condition (peritoneum, ovary, uterus, DIE – deep infiltrating endometrio-\nsis) and one indicating no visible endometriosis (no pathology).\nPurposes Binary as well as multi-label (endometriosis) classification, detec-\ntion and localization tasks with the option of tracking pathology over video\nsegments, thus, augmenting the overall annotated sample count.\nThe remainder of this paper discusses details about the dataset, starting\nwith its creation in Section 2, its structure in Section 3 and, finally, discussing\nits limitations in Section 4 before drawing conclusions in Section 5.\n2 Dataset Creation\nOverall, there are very numerous potential locations for endometriosis 6 and the\ncondition’s size determines its severity level on a scale from one to three. Hence,\nfor creating a a complete dataset in terms of a sufficient amount of examples for\nevery possible combination of location and severity extent requires the collection\nof samples for well over 50 different category types or classes. This, in fact, can\nbe considered an overly challenging task due to the following reasons:\n5 http://www.itec.aau.at/ftp/datasets/GLENDA\n6 e.g. peritoneum, ovary, tube, ligaments, vagina, rectum, bladder, ureter, ...\n\n4 A. Leibetseder et al.\nExpert Knowledge Endometriosis can not reliably be recognized by laymen or\neven untrained medical practitioners, which stresses the need for employing\nspecific experts in the field and at the same time severely reduces the amount\nof capable annotators for such a dataset.\nTime GLENDA has been created with fully active surgeons restricting all an-\nnotation and research effort to non-working hours, which considerably slows\ndown the data collection process.\nRarity Several lesion locations are diagnosed much more rarely than others,\nhence, including them prolongs data accumulation even further.\nCompleteness Finding representative examples of a class for each of the three\npotential severity levels even for a reduced set of categories poses an ad-\nditional challenge outweighing the time and effort for spent annotating, at\nleast for the dataset’s current first version.\nTherefore, for our initial GLENDA version we constrain annotations to a\nsubset of four endometriosis locations (see Section 3 for details) consisting of\nregion-based annotations of single video frames, which either are associated with\nspecific video positions (frame annotations) or sequences over time (keyframe an-\nnotations in video segments). Although for video segments only keyframes are\nannotated, they have been created keeping in mind that all of the endometrio-\nsis regions identified on them are visible throughout the sequences, i.e. camera\nmotion is kept at a minimal level. This offers the possibility of augmenting the\nnumber of annotations by applying annotation tracking mechanisms to these\nsegments (e.g. point/kernel/silhoutte tracking).\n(a) Creating annotations using closed free-\nhand drawings, polygons and rectangles.\n(b) Different endometriosis annotations in\na summary view.\nFig. 2: Dataset creation and exploration using the Endoscopic Concept Annota-\ntion Tool (ECAT).\nThe entire dataset has been created using the Endoscopic Concept Annota-\ntion Tool [6] (ECAT), shown in Figure 2. ECAT is a web-technologies-based tool\nthat allows for importing large video databases and creating concept annotations\nfor video sequences as well as frames. It in particular enables users to annotate\nframes by creating rectangles as well as closed polygons and free-hand drawings,\nmeaning that every annotation always needs to enclose a region. Further, a single\nannotation on a frame that potentially can contain many different annotations, is\n\nGLENDA: Gynecologic Laparoscopy Endometriosis Dataset 5\nrequired to be associated with one of GLENDA’s four endometriosis categories.\nFor video sequences, a keyframe must be chosen first before being able to draw\nan annotation region.\nAs the system can be utilized via a standard web browser, it has been made\nremotely accessible to all involved medical experts for increasing the convenience\nwhen creating annotations. Finally, all data for the current GLENDA version\nhas been collected over the course of four months time.\n3 The GLENDA Dataset\nGLENDA is a multi-faceted endometriosis dataset that has been extracted from\nover 400 gynecologic laparoscopy videos, many of which show endometriosis cases\nof varied severities. It is summarized in Table 1 and composed of following ele-\nments:\nCategories Five categories/classes describe a distinct endometriosis locations:\n– pathology: peritoneum, ovary, uterus, deep infiltrating endometriosis (DIE)\n– no pathology: (no visible endometriosis)\nAnnotations Region-based as well as temporal annotations in the form of:\n– Annotated frames: single video frames annotated with hand-drawn sketches\n(regions), which indicate one or more out of four endometriosis cate-\ngories.\n– Annotated sequences: sets of consecutive video frames associated with\none or several categories over certain periods of time with annotated\nkeyframes (pathology) or no additional region-based annotations (no\npathology).\nAs a consequence of choosing above structure, the dataset allows for a mul-\ntitude of utilization purposes. Splitting up the dataset into pathology and no\npathology images allows for attempting binary classification, disregarding all\nendometriosis sub-classes. Furthermore, when including individual class annota-\ntions multi-class endometriosis prediction can be approached, as already men-\ntioned in above Section 2 potentially by augmenting the amount of annotations\nvia tracking them throughout their corresponding video segments. Aside from\npossible disadvantages outlined in Section 4, additionally collecting video seg-\nments has the advantage of enabling the inclusion of temporal information in\nproposed methodologies for analysis. Finally, the multitude of region-based an-\nnotations can be leveraged for localization tasks as well as representing a basis\nfor learning further annotations.\nFollowing sections more thoroughly describe GLENDA’s class categories as\nwell as structure on a file basis, while pointing out the dataset’s limitations.\n3.1 Categories\nPeritoneum The peritoneum is a serous membrane lining the abdominal cav-\nity (parietal) as well as its contained upper organs (visceral). Endometrial tissue\n\n6 A. Leibetseder et al.\nTable 1: GLENDA summary: number of annotations (annot.) per category (cat.),\nnumber of annotated frames per category, maximum (max.) annotations per\nframe, max. categories (cat.) per frame, number of sequences (seq.) and amount\nof frames.\nCategory* annot. annot.\nframes\nmax. annot.\nper frame seq. frames\nperitoneum 402 203 9 73 6470\novary 51 48 2 15 2478\nuterus 14 8 3 5 475\npathology\nDIE 53 43 3 18 2821\nno pathology 0 0 0 27 13 438\nT otal 520 302 9 (max. cat.: 3) 138 25 682\n*Note: A sequence/keyframe/frame is attributed to a specific category if it is the dominant one in\nall of its corresponding annotations in terms of annotation count/area covered.\n(a)\n (b)\n (c)\n (d)\n (e)\n(f)\n (g)\n (h)\n (i)\n (j)\n(k)\n (l)\n (m)\n (n)\n (o)\nFig. 3: Peritoneum: differing example images (3a - 3e) with corresponding annotations\n(3f - 3j) and video sequence example including keyframe annotations (3k - 3o).\nannotated in GLENDA is associated with the pelvic cavity, which is enclosed\nby the parietal peritoneum. Figure 3 shows various peritoneum dataset exam-\nples together with their annotations 7 (binary images) and selected frames of\n7 Note that due to the possibility of annotating several categories per image, e.g.\nGLENDA includes region-based annotations of up to three classes per image, for\nsimplicity only frames with exactly one associated class have been chosen as exam-\nples.\n\nGLENDA: Gynecologic Laparoscopy Endometriosis Dataset 7\na video sequence with its annotated keyframe (green overlay). Since the peri-\ntoneum covers a very large area and as well surrounds organs that are part of\nother GLENDA classes, corresponding images are often very different to one an-\nother and can contain non-relevant other areas that may even contain additional\nendometrial tissue. Another implication of the membrane’s proportionally large\nsize is its consequently very frequent visibility in many of the dataset’s images,\nbe it pathologoical or non pathological. Visually the peritoneum occurs in a\nmixture of red, yellow and white colors.\n(a)\n (b)\n (c)\n (d)\n (e)\n(f)\n (g)\n (h)\n (i)\n (j)\n(k)\n (l)\n (m)\n (n)\n (o)\nFig. 4: Ovary: differing example images (4a - 4e) with corresponding annotations (4f -\n4j) and video sequence example including keyframe annotations (4k - 4o).\nOvary Apart from carrying several important functions like producing hor-\nmones, the main purpose of the two ovaries is to produce mature ova. Together\nwith the peritoneum class, the ovaries are the most common locations for en-\ndometriosis (see Figure 4 for dataset examples), which is the reason why both of\nthem (together with the fallopian tubes) are the only lesion locations described\nby the rASRM score [2], specifically as well for diagnosingadhesions, i.e. endome-\ntrial tissue connecting other tissue – a class that, however, is not yet included\nin GLENDA. Visually ovaries are easily distinguishable from other organs even\nfor laymen, since their oval-shaped outer capsule in non pathological state is\ncolored in a shade of white contrasting them from the typical red-yellowish color\nspectrum of other anatomical structures.\nUterus The uterus is intended for bringing up a fetus from a fertilized ovum.\nUteri can show different types of endometrial dislocation: endometrial tissue\ngrowing into the muscle wall of the uterus and thickening it is calledadenomyosis\n(adenomyosis may alter the shape and consistency of the uterus), while there is\n\n8 A. Leibetseder et al.\n(a)\n (b)\n (c)\n (d)\n (e)\n(f)\n (g)\n (h)\n (i)\n (j)\n(k)\n (l)\n (m)\n (n)\n (o)\nFig. 5: Uterus: differing example images (5a - 5e) with corresponding annotations (5f\n- 5j) and video sequence example including keyframe annotations (5k - 5o).\nno special term for the case that the tissue is found on the uterine surface, which\nis covered by the visceral peritoneum. Similar to the previous classes, Figure 5\ndepicts various examples for this category together with a sample sequence of\nan affected uterus. A non pathological uterus is pear-shaped, visually appears\nin shades of red and is located in-between the two ovaries and in connection to\nthem via the fallopian tubes.\n(a)\n (b)\n (c)\n (d)\n (e)\n(f)\n (g)\n (h)\n (i)\n (j)\n(k)\n (l)\n (m)\n (n)\n (o)\nFig. 6: Deep Infiltrating Endometriosis (DIE): differing example images (6a - 6e) with\ncorresponding annotations (6f - 6j) and video sequence example including keyframe\nannotations (6k - 6o).\n\nGLENDA: Gynecologic Laparoscopy Endometriosis Dataset 9\nDIE Non-shallow endometriosis that is found on specific locations such as the\nrectum, the rectovaginal space or uterine ligaments is described as Deep Infil-\ntrating Endometriosis (DIE) and is usually rated using the Enzian classification\nsystem [3] in addition to the rASRM score [2]. Figure 6 shows several examples\nof this class including lesions in the pelvic wall and uterine ligaments. Since all\nof the previously described anatomical structures can be affected by DIE ren-\ndering example pictures very similar to the other classes, it is a challenging task\nto classify DIE correctly. Additionally, as this type of endometriosis describes a\nlarge variety of lesion locations, no distinct visual appearance can be attributed\nto this type of class, other than highlighting that typically the color spectrum\nin recorded laparascopic videos lacks green tones.\n(a)\n (b)\n (c)\n (d)\n (e)\n(f)\n (g)\n (h)\n (i)\n (j)\n(k)\n (l)\n (m)\n (n)\n (o)\nFig. 7: No Pathology: differing example images (7a - 7j) and video sequence example\n(7k - 7o).\nNo Pathology Video sequences containing no visible pathology in relation to\nendometriosis are included in the dataset, providing counter examples to above\ncategories. Since this class does not contain any region-based annotations, in ad-\ndition to a sequence showing a non-pathological uterus, Figure 7 particularly in-\ncludes examples of several anatomical structures from above pathological classes\n(e.g. peritoneum and ovaries). Again it is not possible to make any assumptions\nabout the color and shape of objects within this class, since it includes images\ncovering most areas of the pelvic region.\n3.2 Structure\nAlthough originally extracted from videos, GLENDA is an image-based dataset,\nhence, contains a structured collection of images. Besides images of video frames,\n\n10 A. Leibetseder et al.\nall annotations have been extracted separately and are also provided as images,\nalbeit in binary format (as depicted in Figures 3-6, but with the restriction of\none annotation per image). The dataset archive addionally includes a Readme\nfile as well as some dataset statistics in comma separated value (CSV) tables.\nIts directory structure is listed and explained in Figure 8.\nROOT\nDS\nno pathology\nframes\npathology\nannotations\nframes\nReadme.md\nstatisitics\nannotations.csv\nimages.csv\n[DS/pathology/annotations]\nBinary annotation images (1 per annotation) of\nstructure: v VID s FROM-TO/f FID/CLASS a AID.png\n[DS/*pathology/frames]\nVideo frames of structure:\nv VID s FROM-TO/f FID a.jpg\n(optional ” a” indicates that frame has annotation)\n[statistics]\nStatistics such as annotations/images per class, max.\nannotations per image etc.\nFig. 8: GLENDA’s directory structure (placeholders capitalized).\nGLENDA’s folder structure as well as file names reflect all relevant infor-\nmation for utilizing the dataset. In particular every contained video, frame and\nannotation possesses a unique ascending ID number and frames can be mapped\nto their annotations by partial path matching, which can be sped up by first\nremoving all non-annotated frames, i.e. finding frames with an ” a” suffix:\nDS/pathology/frames/v_2401_s_210-318/f_210_a.jpg ->\nDS/pathology/annotations/v_2401_s_210-318/f_210/die_a_629.png\nDS/pathology/annotations/v_2401_s_210-318/f_210/die_a_630.png\nDS/pathology/annotations/v_2401_s_210-318/f_210/die_a_631.png\n...\nRegion-based annotations, which, as mentioned above, are in binary format\n(black background, white annotation) need to appropriately be transformed into\nthe required format, e.g. rectanglar bounding boxes, polygons etc. For conducting\nbinary classification region-based annotations are disregarded, hence, images can\nsimply be retrieved from their corresponding folders (DS/pathology/frames and\nDS/no pathology/frames).\n4 Limitations\nWhen thoroughly examining GLENDA in its current version, several shortcom-\nings can be identified. Although the total amount of images is approximately\n\nGLENDA: Gynecologic Laparoscopy Endometriosis Dataset 11\nbalanced between examples for pathology and no pathology (12K+ vs. 13K+),\nsingle pathology classes are uneven: e.g. peritoneum contains more annotations\nthan all three other classes combined. Especially the sample count of the uterus\ncategory with a total of merely 8 annotated images with 14 annotations is rather\nlow. Therefore, researchers may either choose to omit this class altogether or at-\ntempt to augment it via tracking the annotations over the class’ 5 included video\nsegments. Although this will yield a bigger sample count of 475 images, it is of\ncourse less diverse than creating new annotations.\nPaying even more attention to data diversity, when using the full 25K+ im-\nage corpus with the goal of machine learning based classification, it is important\nto carefully consider data preparation: sequential video frames within small pe-\nriods of time (e.g. a few seconds) are very similar to each other, which can have\na grave detrimental impact on the training process. Thus, training, validation\nand test splits should be constructed from distinct video sequences rather than\nfrom the entire image pool, specifically for non pathological samples, which ex-\nclusively are comprised of merely 27 sequences. Although more effortful, this\napproach has the advantage that the data across the splits is always truly dif-\nferent, preventing the classifier from validating/testing on exactly what it has\nbeen trained on, yielding a perfect classification score. In order to increase the\namount of dissimilar sequences, it may also be feasible to uniformly sample from\nthe existing sequences with a large enough interval or apply a shot detection\ntechnique for finding boundaries between dissimilar content.\nFinally, not least due to the varied sample count for pathological classes,\nin order to ensure an optimal training split that is balanced similarly to the\ndataset, it should be proportional to the amount of examples per class and not\nthe total number of images, i.e. every split should contain a certain percentage\nof examples from each class.\n5 Conclusion\nWith the aim of inspiring research in gynecologic laparoscopy, we introduce\nthe first Gynecologic Laparoscy ENdometriosis DAtaset (GLENDA), created in\ncollaboration with medical experts in the domain of endometriosis treatment.\nGLENDA contains over 25K images, about half of which are pathological, i.e.\nshowing endometriosis, and the other half non-pathological, i.e. containing no\nvisible endometriosis. We thoroughly describe the data collection process, the\ndataset’s properties and structure, while also discussing its limitations. We plan\non continuously extending GLENDA, including the addition of other relevant\ncategories and ultimately lesion severities. Furthermore, we are in the process of\ncollecting specific ”endometriosis suspicion” class annotations in all categories for\ncapturing a common situation among endometriosis experts: at times it proves\ndifficult even for specialists to classify the anomaly without further inspection,\nwhich may be due to several reasons, such as visible video artifacts or DIE regions\nwith a small surface area. Although doubling the amount of classes, having such\ndifficult examples may greatly improve the quality of endometriosis classifiers.\n\n12 A. Leibetseder et al.\nAcknowledgements\nThis work was funded by the FWF Austrian Science Fund under grant P 32010-N38.\nReferences\n1. Andrews, W., Buttram Jr, V., Behrman, S., Carmichael, E., Cohen, M., Dmowski,\nP., Eward, R., Guzick, D., Jones Jr, H., Kistner, R., et al.: Revised american\nfertility society classification of endometriosis: 1985. Fertility and sterility 43(3),\n351–352 (1985)\n2. Canis, M., Donnez, J., Guzick, D., Halme, J., Rock, J., Schenken, R., Vernon, M.:\nRevised american society for reproductive medicine classification of endometriosis:\n1996. Fertility and Sterility67(5), 817–821 (1997). https://doi.org/10.1016/S0015-\n0282(97)81391-X\n3. Keckstein, J.: Endometriosis in the Intestinal Tract – Important Facts\nfor Diagnosis and Therapy. coloproctology 39(2), 121–133 (mar 2017).\nhttps://doi.org/10.1007/s00053-017-0144-5\n4. Leibetseder, A., Petscharnig, S., Primus, M.J., Kletz, S., M¨ unzer, B., Schoeffmann,\nK., Keckstein, J.: Lapgyn4: a dataset for 4 automatic content analysis problems\nin the domain of laparoscopic gynecology. In: Proc. of the 9th ACM Multimedia\nSystems Conf. pp. 357–362. ACM (2018)\n5. Litjens, G., Kooi, T., Bejnordi, B.E., Setio, A.A.A., Ciompi, F., Ghafoorian, M.,\nVan Der Laak, J.A., Van Ginneken, B., S´ anchez, C.I.: A survey on deep learning\nin medical image analysis. Medical image analysis 42, 60–88 (2017)\n6. M¨ unzer, B., Leibetseder, A., Kletz, S., Schoeffmann, K.: Ecat-endoscopic concept\nannotation tool. In: Intl. Conf. on Multimedia Modeling. pp. 571–576. Springer\n(2019)\n7. M¨ unzer, B., Schoeffmann, K., B¨ osz¨ ormenyi, L.: Content-based processing and anal-\nysis of endoscopic images and videos: A survey. Multimedia Tools and Applications\n(jan 2017). https://doi.org/10.1007/s11042-016-4219-z\n8. Schoeffmann, K., Husslein, H., Kletz, S., Petscharnig, S., Muenzer, B., Beecks,\nC.: Video retrieval in laparoscopic video recordings with dynamic content de-\nscriptors. Multimedia Tools and Applications 77(13), 16813–16832 (jul 2018).\nhttps://doi.org/10.1007/s11042-017-5252-2\n9. Stauder, R., Ostler, D., Kranzfelder, M., Koller, S., Feußner, H., Navab, N.:\nThe tum lapchole dataset for the m2cai 2016 workflow challenge. arXiv preprint\narXiv:1610.09278 (2016)\n10. Tsui, C., Klein, R., Garabrant, M.: Minimally invasive surgery: national trends\nin adoption and future directions for hospital strategy. Surgical endoscopy 27(7),\n2253–2257 (2013)\n11. Twinanda, A.P., Shehata, S., Mutter, D., Marescaux, J., de Mathelin, M.,\nPadoy, N.: EndoNet: A Deep Architecture for Recognition Tasks on Laparo-\nscopic Videos. IEEE Transactions on Medical Imaging 36(1), 86–97 (jan 2017).\nhttps://doi.org/10.1109/TMI.2016.2593957\n12. Ye, M., Giannarou, S., Meining, A., Yang, G.Z.: Online tracking and retargeting\nwith applications to optical biopsy in gastrointestinal endoscopic examinations.\nMedical image analysis 30, 144–157 (2016)","source_license":"CC0","license_restricted":false}