If you don't remember your password, you can reset it by entering your email address and clicking the Reset Password button. You will then receive an email that contains a secure link for resetting your password
If the address matches a valid account an email will be sent to __email__ with instructions for resetting your password
Colon polyp detection and optical biopsy are key performance indicators in colonoscopy. Artificial Intelligence (AI) has the potential to greatly improve both. However, practical real-time applications with standard scopes are elusive.
We have published on real-time optical biopsy of diminutive colon polyps using AI, surpassing the 90% negative predictive value (NPV) for adenomas, as per PIVI guidelines. We sought to use latest AI techniques to further improve our optical biopsy performance. In relation to polyp detection, most AI detection tools are trained using still images or videos with obvious polyps. In contrast, we planned our tool around difficult sequences from clinical screening videos that start when the polyp first becomes visible. Finally, an AI model capable of detecting NBI light was the cornerstone that allowed us to propose a “full clinical workflow” for colon polyp detection immediately followed by optical biopsy. Our workflow was optimized to allow for real-time clinical use, a first in this field.
The full workflow captures the video feed from a tower and consists of three distinct AI models: a NBI light detector, a polyp detector, and an optical biopsy. The NBI light detector (Fig. 1a) runs continuously and triggers either the detection mode (white light, Fig. 1b) or the optical biopsy mode (NBI light, Fig. 1c). This allow a seamless interface without the need for a switching signal from either the tower or operator.
The NBI light model was tested on 21,804 unseen frames and achieved a near-perfect accuracy of 99.94%.
The polyp detection model was tested on the polyp approach sequence part of 30 previously unseen colonoscopy videos (>20min each). The model detected polyps with a sensitivity of 79.0% while triggering on 13.7% of frames without polyps. Notably, polyps are detected, on average, 403 milliseconds after their first appearance.
The optical biopsy was tested on videos of 125 previously unseen polyps and achieved a sensitivity of 95.95%, specificity of 91.66%, and NPV of 93.6% (Table 1). Even if the model can abstain when unsure, it committed to a prediction for 97.6% of polyps, an absolute increase of 12.8% over our previous work (Byrne et al. Gut, 2017).
Finally, results are displayed in real-time and the user interface is updated 30 times per second.
We propose the first real-time AI full colonoscopy workflow for automatic detection followed by optical biopsy of colorectal polyps. It consists of three separate AI models allowing for real-time detection of colon polyps, automatic recognition of the switch from white light to NBI, followed by immediate optical biopsy of detected polyps. Detection shows very promising results, especially on difficult approach sequences, and our AI optical biopsy has been even further improved. A clinical trial is planned for the near future.