AUTOMATED MEASUREMENT OF COLONOSCOPY WITHDRAWAL TIME USING CONVOLUTIONAL NEURAL NETWORKS

Aims Withdrawal time (WT) is defined as the time taken from reaching the caecum to exiting the anal canal minus time spent cleaning and performing interventions. Recent years has demonstrated the potential of artificial intelligence (AI) to detect caecal landmarks, however, its ability to detect phases of withdrawal is unexplored.

We aimed to develop convolutional neural networks (CNN) to detect cleaning and interventional phases of withdrawal.

Methods Endoscopy videos were collected prospectively. After the appendicular orifice or ileocaecal valve were first detected, individual frames were annotated with image-level labels.

The first frame an instrument was visualised during polypectomy up until the end of inspecting post-resection margins and biopsies was labelled as ‘intervention’. Frames during suctioning of colonic content or washing were labelled as ‘cleaning’. The remaining frames were labelled ‘withdrawal' frames.

Two ResNet-101 CNNs pre-trained on ImageNet were developed to detect the phases of cleaning and intervention.

Results 87 endoscopy videos and 1,288,319 frames during withdrawal were annotated (Table 1). The procedures were split into training (70%), validation (10%) and testing (~20%) with no overlap of patients.

Evaluated against a test-set of 17 videos, the CNNs identified the interventional frames with 92.4% sensitivity and 95.8% specificity. For cleaning, the sensitivity was 83.0% and specificity 89.5%.

The mean WT was 8:51 (minutes:seconds). The absolute mean error of the AI predicted WT was 39 seconds per procedure. The CNNs correctly categorised 16/17 procedures (94%) as below or above 6 minutes.

Table 1

	Withdrawal frames	Cleaning frames	Intervention frames	Total frames
Frame count	437,359	232,384	618,576	1,288,319

Conclusions We demonstrated the feasibility of CNNs to differentiate the phases of withdrawal to measure WT.

Publication History

Article published online:
14 April 2022