Colonoscopes

Anesthesia was induced by using a mix of 95% O2 and isoflurane 3%, followed by maintenance using a mix of 95%O2 and 1.5–3% isoflurane, all at a rate of 1000ml/min and colonoscopies were performed with a 1.9mm Karl Storz Coloview mini-endoscopic system. The colonoscope was passed into the distal colon under direct endoscopic visualization. Colonoscopies were performed on days 14, 35, and 56 with video documentation.

All patients will receive localization and surgery within 1 week after randomization. Endoscopic tattooing with autologous blood and intraoperative colonoscopy will be performed by two experienced endoscopists who has more than thousands of cases colonoscopies and more than 200 cases of endoscopic mucosal resection or endoscopic submucosal dissection.
For patients who will enroll in autologous blood group, the tattooing will be performed at 24–48 hours before the surgery. When the lesion is identified by endoscopy, the patient’s peripheral venous blood will be collected using a 10 ml simple syringe without heparin preparation. Immediately after blood sampling, 2–3 ml of autologous blood will be injected submucosally at the distal side and proximal side of the lesion (about 2 cm below and above the border of the lesion) using a conventional endoscopic needle without submucosal injection of normal saline. The tattooing with autologous blood will consider to be invisible if both distal and proximal spots was not identified. For those receiving autologous blood localization, the case will be applied intraoperative colonoscopy if the autologous blood tattoo will not be identified or inaccurate in the laparoscopic colectomy.
For patients who will enroll in intraoperative colonoscopy group, the patient will be placed in the modified lithotomy position under general anesthesia with endotracheal intubation. The legs will be opened and positioned in padded stirrups to facilitate the insertion and manipulation of the colonoscope during the operation. After routine laparoscopic exploration, CO₂-insufflated intraoperative colonoscopy will be performed using a flexible videocolonoscope. Upstream small bowel clamping will be applied before intraoperative colonoscopy. During intraoperative colonoscopy, CO₂ pneumoperitoneum will be maintained by the insufflator so that the laparoscope could guide the colonoscope effectively.
After lesion will be identified, a standard laparoscopic colectomy will be performed by two experienced surgeons who has more than 20 years of experience in colorectal surgery with more than 200 cases per year for all enrolled patients. All abdominal operation of laparoscopy will be videotaped. Anastomosis will be performed using the instrumental method. The specimen will be pulled out through a small median incision under the xiphoid (about 3–8 cm).
For those receiving laparoscopic colectomy, the case will be required to be converted to open surgery if one of the following happens: severe or life-threatening intraoperative complications such as intra-abdominal massive haemorrhage, severe organ damage, or other technical or instrumental factors that require a conversion to open surgery.

A clip was defined as a set of $k$ consecutive frames. Clips
were extracted from the colonoscopic videos in a sliding window
fashion with a stride of one to maximize the number of clips. For the
internal dataset, frozen video sequences were excluded to ensure
variation within the clips. Only clips where
$\ge <!--\; \ge \; -->$ 50% of the frames were
labelled as high-quality were included to simulate the clinical setup,
where the endoscopist performs visual diagnosis once a good view of
the polyp is obtained and the polyp features are visible. The
$\ge <!--\; \ge \; -->$ 50% threshold was selected
to ensure a balance between sufficient image quality and the amount of
discarded data. In the Piccolo dataset consecutive frames were not
available as the videos are sampled every 25 frames, so clips were
composed of non-consecutive, ordered frames, and no clips were
discarded due to image quality.
The LRCN model was trained with each clip as an input sample, whereas
the ConvNet was trained with all the frames included within the LRCN
clips, ensuring the same frames were utilised for both methods,
although less samples were used for LRCN. After excluding white light
sequences, low-quality clips and lesions with less than k frames, the
baseline model was trained with a total of 27,087 frames from 197
polyps from 89 colonoscopy procedures, for
$k=15$ . As a note, Fig. 3 shows an example of how clips were
discarded when they contained non-annotated frames, reducing the
amount of available samples.
Random sampling of 5000 frames was performed on each epoch, re-sampling
each time, to minimise overfitting [41 ]. Data augmentation was applied in such a way as to
guarantee identical augmentations within clips. The augmentation
operations consisted of random affine transformations (rotation,
translation and scaling) and random colour transformations
(brightness, contrast and saturation). Finally, the images were
preprocessed by cropping around the polyp boxes annotated by experts,
followed by resizing the images to 224 by 224 pixels and an intensity
normalization step. Only the polyp area was used as an input to the
networks, discarding the remainder of the image. This ensures that the
adenoma classification model can be used in a clinical setting, where
more than one polyp can be present in a frame.
All models were trained with 5-fold patient cross-validation. For all
experiments, the same folds were respected, to ensure a fair
comparison between models. For each fold, each patient’s video was
used for either training or testing following an 80-20% split,
avoiding any data contamination. The patient splits were generated
optimizing the distribution balance between the training and testing
sets in terms of the number of NBI polyp frames, the number of
different lesions, the polyp size (in pixels), the polyp types and the
quality of the images.