Colonoscopy Coverage Revisited: Identifying Scanning Gaps in Real-Time

TL;DR

This work addresses missed polyps during colonoscopy by tackling deficient local coverage in real time, especially during poor-visibility intervals where dense $3D$ reconstructions struggle. It introduces a two-phase approach: Phase I uses self-supervised learning (simCLR) to classify visibility and identify good-visibility segments and gaps, achieving high per-frame accuracy ($0.94$); Phase II learns low-dimensional scene descriptors via contrastive learning to decide whether before-and-after-gap scenes are the same, using automatically generated training data from optical-flow-based endoscope motion. Evaluations on 250 colonoscopy procedures with clinician annotations report a gap-classification ROC AUC of $0.90$ and sensitivity $0.75$ with specificity $0.90$, demonstrating practical real-time alerting to revisit areas with potential loss of coverage. The approach complements existing $3D$-reconstruction methods and offers a pathway toward real-time guidance and bookmarking of regions of interest in the colon.

Abstract

Colonoscopy is the most widely used medical technique for preventing Colorectal Cancer, by detecting and removing polyps before they become malignant. Recent studies show that around one quarter of the existing polyps are routinely missed. While some of these do appear in the endoscopist's field of view, others are missed due to a partial coverage of the colon. The task of detecting and marking unseen regions of the colon has been addressed in recent work, where the common approach is based on dense 3D reconstruction, which proves to be challenging due to lack of 3D ground truth and periods with poor visual content. In this paper we propose a novel and complementary method to detect deficient local coverage in real-time for video segments where a reliable 3D reconstruction is impossible. Our method aims to identify skips along the colon caused by a drifted position of the endoscope during poor visibility time intervals. The proposed solution consists of two phases. During the first, time segments with good visibility of the colon and gaps between them are identified. During the second phase, a trained model operates on each gap, answering the question: Do you observe the same scene before and after the gap? If the answer is negative, the endoscopist is alerted and can be directed to the appropriate area in real-time. The second phase model is trained using a contrastive loss based on auto-generated examples. Our method evaluation on a dataset of 250 procedures annotated by trained physicians provides sensitivity of 0.75 with specificity of 0.9.

Figures (5)

• Figure 1: Our solution starts by detecting time segments with good visibility of the colon and gaps between them. For each such gap we answer the question: Do you observe different scenes before and after the gap? If the answer is positive, the endoscopist is alerted to revisit the area in real-time.
• Figure 2: To achieve high accuracy visibility classifier, we train an encoder in an unsupervised manner and then train a binary classifier the resulting embeddings using the labeled data. Further improvement is made by smoothing predictions based on similarity distances, resulting in $94\%$ accuracy on the validation set.
• Figure 3: Endoscope displacement estimation is based on optical-flow calculation between consecutive frames using the amount of flow trough the frame boundary (see kelner2022motion).
• Figure 4: We simulate random artificial gaps of various duration in good-visibility video segments, estimate the endoscope motion within these simulated gaps, and get this way reliable training examples for our overall task. Gaps associated with low accumulated motion contribute a 'same-scene' training example ($c_k=1$), while high-motion gaps refer to a different scene pair ($c_k=0$).
• Figure 5: Direct gap classification: ROC curve evaluated on the whole dataset (750 gaps).