Mieke Bus

110 Chapter 7 Emerging techniques that enhance tumour visualisation with OCT can, therefore, potentially reduce unnecessary biopsies, without extra morbidity. This scenario is especially interesting for the case of cytology suspicious of high-grade malignancy and negative ureterorenos- copy findings, when so-called at-random biopsies are taken to exclude or confirm carci- noma in situ (CIS). OCT measurements via the working channel of the ureterorenoscope could be of value in the diagnosis of CIS, i.e. guiding the biopsy procedure by visualizing CIS lesions. Nowadays, CIS detection knows several challenges and OCT confirmation is limited by reliance on white light URS to identify the suspicious lesions. Combining NBI-URS with OCT can potentially overcome this limitation. As we demonstrated in this thesis, CIS can be recognized on OCT images as a thickened urothelial layer. However, there are more scenar- ios that can cause a thickened urothelial layer. To dilate the ureter, some patients receive a double J-catheter pre-operatively. This catheter can be irritating for the urothelial layer, causing oedema. This urothelial oedema might be seen as a thickened layer on OCT imaging and it is difficult to distinguish double J-catheter effect from CIS on OCT images. In addition, OCT is known for false-positives in inflammatory states as inflammation can cause a thick- ened urothelial layer, resembling UTUC lesions on OCT images. (22) For this reason, biopsies will remain, for now, necessary to confirm or exclude urothelial cancer in the upper urinary tract. At this moment, OCT has to be seen as a valuable addition to the diagnostic work-up in UTUC and not as a replacement for histology. Finally, what is perhaps the most exciting possible feature of OCT is a combination of endos- copy, artificial intelligence (AI) with OCT. AI can be explained as a machine intelligence sim- ilar to natural intelligence displayed by humans, including cognitive functions as learning and problem solving. Deep learning (DL) is a method that autonomously learns features and tasks from a training dataset, for example medical images. Medical image-based diagnoses such as pathology radiology and endoscopy are expected in the very near future to be the first in the medical field using AI. (23) In gastroenterology this technique is already able to automatically detect cancer or other pathological endoscopic findings during endoscopy and is expected to be mainstream technology in the next few decades by helping endos- copists by providing a more accurate diagnosis by automatically detecting and classifying endoscopic lesions. (23, 24) This technique can be of great use in cystoscopy and ureterorenos- copy as well for detecting and classifying pathological findings. One of the most important factors for the development of DL is the availability of large amounts of high-quality endo- scopic images for DL training dataset. In OCT imaging large volumes of clinical images are obtained, making it an excellent target for AI modalities. (25) AI using DL, based on endoscopic images combined with OCT datasets gives the urologist of the future optimal quantative information of pathology in the urinary tract.