Semi-supervised Quality Evaluation of Colonoscopy Procedures (2305.10090v1)

Abstract: Colonoscopy is the standard of care technique for detecting and removing polyps for the prevention of colorectal cancer. Nevertheless, gastroenterologists (GI) routinely miss approximately 25% of polyps during colonoscopies. These misses are highly operator dependent, influenced by the physician skills, experience, vigilance, and fatigue. Standard quality metrics, such as Withdrawal Time or Cecal Intubation Rate, have been shown to be well correlated with Adenoma Detection Rate (ADR). However, those metrics are limited in their ability to assess the quality of a specific procedure, and they do not address quality aspects related to the style or technique of the examination. In this work we design novel online and offline quality metrics, based on visual appearance quality criteria learned by an ML model in an unsupervised way. Furthermore, we evaluate the likelihood of detecting an existing polyp as a function of quality and use it to demonstrate high correlation of the proposed metric to polyp detection sensitivity. The proposed online quality metric can be used to provide real time quality feedback to the performing GI. By integrating the local metric over the withdrawal phase, we build a global, offline quality metric, which is shown to be highly correlated to the standard Polyp Per Colonoscopy (PPC) quality metric.

• The paper introduces a novel semi-supervised framework that leverages contrastive self-supervised learning to extract visual embeddings for quality evaluation.
• It employs k-means clustering on learned features to compute online and offline quality metrics that correlate with polyp detection sensitivity.
• The study demonstrates that procedures with higher quality metrics significantly increase the likelihood of polyp detection, suggesting potential for real-time feedback improvement.

The paper introduces a novel approach to semi-supervised quality evaluation of colonoscopy procedures, aiming to address the challenge of missed polyps during colonoscopies, which is often attributed to operator-dependent factors. The work proposes both online and offline quality metrics derived from the visual data of colonoscopies, leveraging machine learning models to assess the quality of the procedure.

Key aspects of the methodology include:

• Learning visual representations of colonoscopy video frames using contrastive self-supervised learning, specifically the SimCLR framework. The loss function is defined as:

  $$\mathcal{\ell} (z_i^1,z_i^2) = -log \frac{exp(sim(z_i^1,z_i^2)/\tau)}{\sum_{k \ne i} \sum_{a=1}^2 \sum_{b=1}^2 exp(sim(z_i^a,z_k^b)/\tau)}$$

  where:

  • $z_i^1$ and $z_i^2$ are the embedding vectors of two augmented versions of frame $x_i$
  • $$sim(u, v) = u^T v / \lVert u \rVert \lVert v \rVert$$ is the cosine similarity
  • $\tau$ is a temperature parameter
• Applying cluster analysis on the learned representations using $k$-means to identify semantically meaningful groups of frames. Instead of hard assignments, the probability of the $i$-th frame belonging to the $k$-th cluster is computed as:

  $$r_{i,k} = Prob(f_\theta(x_i) \in k) \sim \left[\frac{1}{\|f_\theta(x_i) - c_k \|_2^2}\right]^\alpha$$

  where:

  • $f_\theta(x_i)$ is the feature representation of frame $x_i$
  • $c_k$ is the center of the $k$-th cluster
  • $\alpha=16$ is a constant
• Developing an online quality metric, $Q$, that predicts the likelihood of detecting a polyp based on the visual appearance of video segments. The classifier $Q(\cdot)$ is trained to map the cluster assignments $\{\overline{r_{i,k}\}_{k=1}^K$ to the detection of a polyp in the subsequent 2 seconds, achieving a polyp detection prediction accuracy of 64% on the test set.
• Formulating a method to estimate the chance of detecting a polyp at a given time $t$, denoted as $P(D|E,Q)$, which represents the probability of detecting a polyp ($D$) given its existence ($E$) and the quality of the procedure ($Q$). This is approximated using Bayes' rule and empirical estimation of $P(Q)$ and $P(Q|D)$.
• Creating an offline quality metric, $Q_{\text{Offline}}$, by integrating the online metric $Q$ over the entire withdrawal phase of the colonoscopy, according to the equation:

  $Q_{\text{Offline} = \sum _{i \in \text{withdrawal} Q \left(\{r_{i,k}\}_{k=1}^K\right)$

The experimental results demonstrate a strong correlation between the proposed online quality metric $Q$ and the polyp detection sensitivity (PDS). The likelihood of detecting an existing polyp, $P(D|E,Q)$, increases with higher values of $Q$. Furthermore, the offline quality metric $Q_{\text{Offline}}$ shows a high correlation with the standard Polyps Per Colonoscopy (PPC) metric. Procedures with higher $Q_{\text{Offline}}$ values are more likely to have detected polyps.

The paper suggests that real-time feedback based on the online quality metric could potentially improve the quality of colonoscopy procedures and increase the adenoma detection rate (ADR).