Összesen 1 találat.


001-es BibID:BIBFORM071334
Első szerző:Török Péter (szülész-nőgyógyász)
Cím:Differentiating tissues and organs in endoscopic images using a convolutional neural network / Török Péter, Lampé Rudolf, Harangi Balázs, Lőrincz Judit
Megjegyzések:TitleDifferentiating tissues and organs in endoscopic images using a convolutional neural networkStudy questionIs it possible to identify different tissues and organs during endoscopy supported by a software.Summary answer During the learning curve of laparoscopy trainees can get help in differentiating tissues by an automated digital image processing based decision support systemWhat is known already Endoscopy surgery is the part of everyday work of surgeon. In an experienced hand provides less postoperative morbidity, complications and the time periods of hospital stay and returning to normal activity, so it has a lot of advantages compared to the laparotomy.Opposing to open-surgery, the recognition of organs is rather difficult in laparoscopy because of the lack of the tactile information. Knowing detailed anatomy, having experience letting us to recognize structures in the abdominal cavity. Because of the anatomical variations, tissue identification is not always sure relying on the visual information.Study design, size, duration COur dataset has been collected retrospectively during 35 different gynecological endoscopic operations at the Department of Obstetrics and Gynecology of the University of Debrecen. The videos have been recorded by a high definition 1-MOS endoscopic camera at 30 frames/sec rate and resolution of 1920?1080 pixels.Participants/materials, setting, methods Data of patients scheduled for gynecological endoscopic operations are analyzed. The medical expert or an assistant should manually mark the region of interest. Then, the maximum number of sub-images of size 224?224 pixels are cut off along the axis from the video frame. Finally, the classification problem is solved automatically using a convolutional neural network with the resulted labels are pinned on the corresponding organs in the video frame.Main results and the role of chance We have presented an approach to develop an application, which helps medical experts with performing endoscopic surgeries. Our effort primarily addressed the drawback of losing tactile information during key-hole surgery in the recognition of different organs. To address this problem, we have developed a semi-automatic tool, which requires a manual annotation regarding the axis of the interested organs first. Then, several sub-images covering the selected organs are extracted and classified by a fine-tuned GoogLeNet convolutional neural network.The classification performance of the fine-tuned GoogLeNet model on our test dataset considering the top-1 error rate is 0.193 at sub-image level. That is, 403 out of the 500 test images have been classified correctly. However, notice that these sub-images are only small, non-overlapping segments of the interested organs. That is, it is reasonable to fuse these label information for recognizing the corresponding organ. To do so, we have applied the simple majority-voting rule on the 4-5 labels supplied by the sub-images for the same organ. In this way, our proposed approach has reached 94.2% final accuracy regarding this binary classification task.Limitations, reasons for cautionOur collected dataset is relatively small with containing insufficient number of images to train a complex neural network, so we should extend the size of our dataset.Wider implications of the findings Using the software made by the results of the study, accuracy of the tissue/organ recognition could be increased during training laparoscopic technique, or for the experts as well.Study funding/competing interest(s) This work was supported in part by the projects GINOP-2.1.1-15-2015-00376 and VKSZ 14-1-2015-0072, SCOPIA: Development of diagnostic tools based on endoscope technology supported by the European Union, co-financed by the European Social Fund.
Tárgyszavak:Orvostudományok Klinikai orvostudományok idézhető absztrakt
konvolúcionális neurális hálózat
Megjelenés:Human Reproduction 32 : suppl. 1 (2017), p. 487. -
További szerzők:Lampé Rudolf (1983-) (szülész-nőgyógyász) Harangi Balázs (1986-) (programtervező matematikus) Lőrincz Judit (1988-) (általános orvos)
Internet cím:DOI
Intézményi repozitóriumban (DEA) tárolt változat
Rekordok letöltése1