Introduction
The use of optical diagnosis for “resect and discard” and “diagnose and leave” strategies has been postulated for several years, but despite promising early results this has not been implemented outside of expert centres due to the challenge of accurate characterisation and sizing (1,2). Computer aided diagnosis (CADx) has shown promising potential for improving optical diagnosis. However, there is still no data on real-time sizing and characterisation together which is what will be required for the introduction of this strategy and meeting Preservation and Incorporation of Valuable endoscopic Innovations (PIVI) thresholds.
Here we aimed to assess the performance of a novel CADx algorithm and its use during real-time colonoscopy.
Aims & Methods
A dedicated colon CADx algorithm was developed to categorise polyps by histology (neoplastic or non-neoplastic) and by size (diminutive or non-diminutive). In the first phase of this study, the system was tested on a customised video based platform using prospectively collected endoscopy videos. Ground truth for characterisation was histology and for sizing was the use of an instrument e.g. biopsy forceps along with 3 expert endoscopists review.
The next phase was a pragmatic study to test real-time performance of CADx during colonoscopy. In this phase endoscopists were asked to characterise and size of polyps which were removed or biopsied for histology. Videos were recorded for external validation by 3 expert endoscopists.
Results
In video based analysis, 292 polyp videos (202 neoplastic and 90 non-neoplastic) were assessed using the CADx device. Video based AI sizing demonstrated sensitivity, specificity, accuracy and NPV for non-diminutive size of 93.40%, 79.03%, 93.40% and 95.45% respectively. Video based characterisation results are summarised in Table 1.
In real-time analysis, 116 (66 neoplastic and 50 non-neoplastic) polyps in 36 patients were assessed with CADx during real-time colonoscopy. Real-time characterisation results are summarised in Table 1. Real-time concordance for sizing between endoscopist and AI as diminutive or non-diminutive was 84.61%. Subgroup analysis was carried out for video based and real-time CADx for performance on diminutive polyps. The NPV for diminutive rectosigmoid polyps for adenomatous diagnosis was 93.3% in real-time.
There was 97.2% concordance between surveillance intervals (ASGE and BSG guidelines) based on histology versus AI diagnosis for diminutive polyps and histology for non-diminutive.
| Table 1. Overall CADx performance |
| | Video based CADx (n=292) | Real-time CADx (n=116) |
| | WLI | IE | WLI | IE |
| Sensitivity (%) | 86.14 | 91.76 | 89.20 | 90.50 |
| Specificity (%) | 85.56 | 78.75 | 82.70 | 94.10 |
| Accuracy (%) | 85.96 | 87.79 | 87.10 | 92.10 |
| NPV (%) | 73.33 | 80.77 | 86.00 | 88.90 |
Conclusion
This is the first study demonstrating the outcome of AI based sizing and characterisation of polyps in real time. The system meets PIVI 1 and PIVI 2 threshold suggesting we are close to being able to implement ‘resect and discard’ and ‘diagnose and leave strategies’. Further validation from a multicentre prospective study will help guide this.
References
1. Ignjatovic A, East JE, Suzuki N, Vance M, Guenther T, Saunders BP. Optical diagnosis of small colorectal polyps at routine colonoscopy (Detect InSpect ChAracterise Resect and Discard; DISCARD trial): a prospective cohort study. Lancet Oncol. 2009 Dec;10(12):1171-8.
2. Rees CJ, Rajasekhar PT, Wilson A, Close H, Rutter MD, Saunders BP, East JE, Maier R, Moorghen M, Muhammad U, Hancock H, Jayaprakash A, MacDonald C, Ramadas A, Dhar A, Mason JM. Narrow band imaging optical diagnosis of small colorectal polyps in routine clinical practice: the Detect Inspect Characterise Resect and Discard 2 (DISCARD 2) study. Gut. 2017 May;66(5):887-895.
Disclosure
Professor Bhandari has received research grants or is the advisory board for Fujifilm, Boston, Olympus, Pentax, 3-D matrix, NEC (Japan), Medtronic