Optical Diagnosis of Neoplasia Using Artificial Intelligence

Computer-Aided Diagnosis and cognitive forcing Computer-aided diagnosis (CADx) for colonoscopy is expected to improve physicians' ability to predict colorectal polyp pathologies (optical diagnosis). However, recent randomized trials indicate that collaboration between humans and CADx yields lower performance than CADx alone.

This suggests suboptimal human-AI interaction due to users' under-reliance on CADx advice, favoring their inherent biases over critically assessing AI suggestions. Psychological interventions, such as cognitive forcing, aim to address this by encouraging crucial assessment of AI suggestions. One example, delayed display of CADx suggestions may promote proactive thinking by giving endoscopists enough time to consider polyp pathologies before receiving CADx suggestions, potentially leading to critical and optimal human-AI interaction. The effectiveness of such cognitive forcing was observed in experimental studies of AI for mammography reading.

However, despite its potential, no studies have evaluated the impact of such interventions on optical diagnosis accuracy in colonoscopy. To investigate the value of psychological intervention in optical diagnosis, the investigators will conduct an ex-vivo randomised controlled study.

Study aim:

This is an ex-vivo randomised controlled study. The hypothesis of our study is that cognitive forcing by delaying display of CADx suggestions facilitates physicians' critical thinking, leading to better clinical outcomes in optical diagnosis in colonoscopy.

Comercially CADx device and video details:

In this study, the investigators are going to use the comercially available CADx system (GI Genius CADx, manufactured by Cosmo Intelligent Medical Devices and distributed by Medtronic Corp). The GI Genius CADx highlights the suspicious area for polyps on-screen with bounding boxes and provides optical diagnosis prediction (i.e. adenoma, non-adenoma). All the endoscopists will be given the information that the CADx tool used in the study is GI Genius with a link to their product overview.

The investigators will collect 100 colonoscopy videos of 100 different diminutive polyps from the Polyp Image BAnk database (PIBAdb). In accordance with the real-world prevalence, 65 polyps will be neoplastic while the remaining 35 will be non-neoplastic. The duration of each video will be adjusted to contain 15 second appearance of the lesion including wite light (WL) and narrow band imaging (NBI). All the polyps should be \< = 5 mm. The original videos were recorded without having any CADx interaction.

PIBAdb contains 507 videos of diminutive colorectal polyps with WL and NBI, of which 231 have a duration of 15 seconds or more. All the videos contain polyps with their histopathology available (i.e. adenoma, sessile serrated lesions, traditional serrated adenomas, invasive, hyperplastic and non-neoplastic). The polyps that have no histology category are going to be excluded from the study.

The investigators will use this database as a pool to select the 100 study videos. First, the investigators are going to split the pool of data into two groups according to polyp histology: one pool containing neoplasia (i.e. adenoma, sessile serrated lesions and traditional serrated adenoma) and the other containing non-neoplasia (i.e. hyperplastic and non-epithelial neoplastic). In each pool, 20 videos will be randomly selected as a first step. These 20 videos will be assessed if they meet our inclusion and exclusion criteria (see below), leaving only eligible videos. This selection process will be repeatedly done until the investigators collect 65 videos of polyps with neoplasia and 35 videos of polyps with non-neoplasia. The figure below shows how the video selection process takes place.

GI Genius´s specification The GI Genius system processes the videos at 50-60 frames per second and in high-definition format to have a good quality of the videos. To process the videos, the GI Genius splits them into two different streams by dedicated video card. One stream is transmitted to the first path to the output without any processing of the AI algorithm. The second stream is sent to the AI algorithm (i.e. second path) and after appropriate computation, if there is a polyp present on-screen appears an overlay in the output highlighting the polyp. The system was designed to work on unaltered WL video streams, but it works under both WL and NBI similarly.The transmission of the video stream to GI Genius is through a Serial Digital Interface (SDI) cable, acquiring the output stream from the video displaying in a computer. Finally, the SDI output stream is transmitted to the monitor containing the original video stream with additional markers superimposed on it.

How to make CADx overlaid videos In the present study, the selected videos of colorectal lesions will be stored in our high-spec computer system first. Then, these videos will be transmitted to GI Genius with an SDI cable. A capture card (DeckLink 8K Pro Mini, Blackmagic) is integrated into the computer system that converts the recorded video output to SDI signal, which allows GI Genius to process the transmitted videos.

All 100 selected videos will be processed by GI Genius in two different ways. In the first set, only the frames from the last 3 seconds of each 15-second video will be processed, and CADx suggestions will appear only in that final 3-second segment. This first set will be shown to endoscopists who are allocated to the intervention arm. In the second set, all frames from the entire 15-second duration will be processed, and CADx suggestions will be overlaid on every frame. This first set will be shown to endoscopists who are allocated to the control arm.

• The reliance level on artificial intelligence, measured using the C value of the signal detection theory [Срок оценки: Through study completion, an average of 1 year]
• Discrimination (d´) level of neoplastic lesions based on the signal detection theory [Срок оценки: Through study completion, an average of 1 year]
• Receiver Operating characteristic (ROC) curve to determine overall discrimination in the signal detection theory [Срок оценки: Through study completion, an average of 1 year]
• Proportion of high confidence diagnosis [Срок оценки: Through study completion, an average of 1 year]
• Association between the reliance level (C value) on AI and endoscopists age,sex, level of expertise in colonoscopy or CADx, confidence level and area of procedence [Срок оценки: Through study completion, an average of 1 year]
• Specificity of the optical diagnosis for neoplastic lesions. [Срок оценки: Through study completion, an average of 1 year]
• Positive and negative predictive values and accuracy of the optical diagnosis for neoplastic lesions. [Срок оценки: Through study completion, an average of 1 year]
• Sensitivity of the optical diagnosis for neoplastic lesions by endoscopists' age, sex, level of expertise in colonoscopy and CADx and area of procedence. [Срок оценки: Through study completion, an average of 1 year]
• Specificity of the optical diagnosis for neoplastic lesions by endoscopists' age, sex, level of expertise in colonoscopy and CADx and area of procedence. [Срок оценки: Through study completion, an average of 1 year]
• Positive and negative predictive values and accuracy of the optical diagnosis for neoplastic lesions by endoscopists' age, sex, level of expertise in colonoscopy and CADx and area of procedence. [Срок оценки: Through study completion, an average of 1 year]
• Association between endoscopists' sensitivity in optical diagnosis of neoplasia and their reliance level on CADx.suggestions [Срок оценки: Through study completion, an average of 1 year]
• Survey responses. [Срок оценки: Through study completion, an average of 1 year]