A120 AUTOMATED BOWEL PREPARATION DETECTION WITH DEEP. CONVOLUTIONAL NEURAL NETWORKS

Background

Introduction: Bowel preparation inadequacy has been shown to increase post-colonoscopy colorectal cancer. As such, the USMSTF recommends repeating colonoscopy within 1 year if bowel preparation is inadequate. However, bowel preparation documentation is variable in clinical practice, and physician recommendations adherent to USMTF guidelines are inconsistent.

Aims

Aims: We present an automated computer assisted method using deep convolutional neural networks to detect bowel preparation and adequacy of bowel preparation with the Boston Bowel Preparation Scale (BBPS).

Methods

Methods: We extracted 38523 images of colonic lumen between 2015 and 2017 from screening colonoscopies. Bowel preparation scores were assessed with BBPS. Adequate bowel preparation was defined as BBPS ≥2, and inadequate bowel preparation was defined as BBPS <2. The dataset was split into 26966 images for training, 7704 for validation, and 3853 for testing. Training data was sampled with replacement from a multinomial distribution to balance subclass distributions in each batch. We developed 2 convoluted neural networks (CNN) using PyTorch with a Densenet-169 backbone pre-trained on ImageNet and fine-tuned on our data for classifying adequacy of bowel preparation (binary) and for subclassification of BBPS (multi-class). We used Adam optimiser with an initial learning rate of 3x10^-4 and a scheduler to decay the learning rate of each parameter group by 0.1 every 7 epochs along with focal loss as our criterion for both classifiers.

Results

Results: The overall accuracy on the test data set for BBPS subclassification was 0.91. The sensitivity for BBPS 0, 1, 2 and 3 were 0.84, 0.91, 0.86, and 0.96, respectively. The specificity for BBPS 0, 1, 2, and 3 were 1.00, 0.98, 0.95, and 0.93, respectively. The overall accuracy of the test data set for adequacy of bowel preparation was 0.97. The sensitivity for adequacy of bowel preparation for BBPS <2 and BBPS ≥2 was 0.92 and 0.99, respectively. The specificity for adequacy of bowel preparation for BBPS <2 and BBPS ≥2 was 0.99 and 0.92, respectively.

Conclusions

Conclusion: We present an automated computer-assisted detection method of bowel preparation with deep convolutional neural networks. The algorithm is capable of accurate classification of adequacy of bowel preparation (97%) and subclassification of bowel preparation (91%) with high sensitivity and specificity. This algorithm can be applied to automate documentation of bowel preparation and adequacy of bowel preparation. Additional studies will need to be conducted to demonstrate its applicability in real-time colonoscopy.

Funding Agencies

None

A137 LAPAROSCOPIC ENDOSCOPIC COOPERATIVE SURGERY FOR GASTROINTESTINAL STROMAL TUMOURS: EARLY EXPERIENCE IN A CANADIAN ACADEMIC CENTRE

Background

Surgical resection with laparoscopic gastric wedge resection is commonly conducted for local management of gastrointestinal stromal tumours (GIST). However, resection margins are often difficult to appreciate for lesions with larger endophytic components. As a result, tumour margins may be compromised or excess tissue resected. Laparoscopic endoscopic cooperative surgery (LECS) was developed in Japan to overcome these technical challenges in the resection of subepithelial lesions, including GISTs. Here, we present a case report of an early Canadian experience utilizing LECS in the management of gastric GIST.

Aims

To describe a case report of an early Canadian experience of LECS for the resection of a gastric GIST.

Methods

We performed a review of the literature and describe a case of LECS.

Results

We present a 70-year-old female referred to our centre for endoscopic resection of a 2.5x2.5cm histologically confirmed gastric GIST (low mitotic index and no known metastases). Repeat endoscopic evaluation at our centre confirmed a 25-30mm subepithelial lesion with both exophytic (small) and endophytic (large) components. After tumour board review, we opted for a LECS approach. In the OR, the lesion was identified endoscopically and marked with a Dual J-Knife (Olympus). The margins were injected with a combination of Voluven, methylene blue, and dilute epinephrine. A circumferential incision was then completed using standard ESD technique. The lesion was subsequently identified laparoscopically, with endoscopic guidance, along the lesser curvature. The lesser omentum was mobilized for clear visualization of the serosa around the lesion. A full thickness incision was made endoscopically along the distal aspect of lesion. Full thickness resection was continued endoscopically for one third of the circumference of the lesion until gastric insufflation became compromised. Full thickness resection was completed laparoscopically under endoscopic guidance with grossly negative margins. The defect was closed with running laparoscopic sutures. Endoscopic leak test was performed which was negative. The specimen was retrieved and follow up pathology demonstrated a GIST with low mitotic index and negative margins without tumour rupture.

Conclusions

In a review of the literature, LECS appears to minimize tissue resection while maintaining R0 resection rates. This technique is especially useful for subepithelial lesions with larger endophytic and transmural components. It has an excellent safety profile with a less than 5% anastomotic leak rate. As such, the literature supports LECS as a suitable procedure for gastric subepithelial lesions &lt;50 mm. However, further studies are needed to compare it systematically to conventional laparoscopic wedge resection in addition to other innovative endoscopic techniques such as STER and EFTR.

Funding Agencies

None