Near Focus NBI-Driven Artificial Intelligence for the Diagnosis of Gastro-Oesophageal Reflux Disease
Near Focus NBI-Driven Artificial Intelligence for the Diagnosis of Gastro-Oesophageal Reflux Disease
Gastro-oesophageal reflux disease (GORD) is a chronic condition with symptoms arising secondary to the reflux of stomach contents (Vakil et al., 2006). It is divided into four phenotypes: Erosive Oesophagitis (EO), Non-Erosive Reflux Disease (NERD), Reflux Hypersensitivity (RH), Functional Heartburn (FH) (Nikaki, Woodland and Sifrim, 2016). The definition of these phenotype have evolved with the addition of diagnostic tests and methods of their interpretation, the most recent being the Lyon Consensus Statement (Gyawali et al., 2018). The majority of patients presenting with symptoms suggestive of GORD have no mucosal lesion seen at endoscopy (Nikaki, Woodland and Sifrim, 2016). Studies have shown a relation of increased IPCL numbers with GORD. This study aims to build a fully autmoated AI model using Near-Focus NBI images on patients with symptoms suggestive of GORD phenotyped in accordance with the Lyon Consensus.
Gastro-oesophageal reflux disease (GORD) is a chronic condition with symptoms arising secondary to the reflux of stomach contents (Vakil et al., 2006). It is divided into four phenotypes: Erosive Oesophagitis (EO), Non-Erosive Reflux Disease (NERD), Reflux Hypersensitivity (RH), Functional Heartburn (FH) (Nikaki, Woodland and Sifrim, 2016). The definition of these phenotype have evolved with the addition of diagnostic tests and methods of their interpretation, the most recent being the Lyon Consensus Statement (Gyawali et al., 2018). The majority of patients presenting with symptoms suggestive of GORD have no mucosal lesion seen at endoscopy (Nikaki, Woodland and Sifrim, 2016). Complications of GORD, such as peptic stricture and Barrett's oesophagus, can be readily diagnosed using WLE. The diagnosis of reflux oesophagitis with standard-definition WLE is well described in the Los Angeles (LA) classification (Armstrong et al., 1996) and validated (Lundell et al., 1999) with LA grades C and D confirmatory of GORD (Gyawali et al., 2018). Furthermore, AET is demonstrated to increase with LA classification A to D (Lundell et al., 1999). There is, however, only a modest inter-observer agreement between LA grades (Kappa coefficient 0.4), especially A and B. Furthermore LA grade A oesophagitis is detected in up to 17.0% of asymptomatic patients (Nozu and Komiyama, 2008; Zagari et al., 2008). A diagnosis of GORD and decisions for anti-reflux surgery cannot be made on this basis, mandating pH testing to confirm GORD.
Narrow Band Imaging
Imaging of the gastro-oesophageal junction using high definition Olympus H260 scope using the LA classification of GORD with WLE and NBI demonstrated improvement in overall interobserver reproducibility when used in a combination compared with WLE alone; k 0.62 vs 0.45 (<0.05)(Lee et al., 2007). Features identified using digital magnification NBI at the squamo-columnar junction in cases of EO (n=41; LA grade A and B), NERD (n=36) and controls (n=32) include micro-erosions (100% EO; 52.8% NERD; 23.3% controls), increased vascularity (95.1% EO; 91.7% NERD; 36.7% controls) and round pit patterns (4.9% EO; 5.6% NERD; 70% controls). Increased vascularity combined with absence of round pit pattern distinguishes NERD from controls with sensitivity and specificity 86.1% and 83.3%. Inter-observer agreement in this single centre study was good for increased vascularity (k=0.95) and micro erosions (k=0.89) but low for pit pattern (k=0.59) (Fock et al., 2009).
Intra-papillary capillary loops (IPCLs) are mucosal capillaries arising from the submucosal vein to the papilla, usually arranged in a regular 'dot' like fashion approximately 100micrometres apart (Inoue, 2001). The visualisation of oesophageal IPCLs with NBI is well documented and form the basis of a NBI classification for squamous neoplasia (Inoue et al., 2015). IPCL morphology changes have been proposed in patients being investigated for NERD, in particular dilatation and elongation of IPCLs in patients with NERD with magnification NBI (Kato et al., 2006).
NBI with optical magnification for the diagnosis of GORD has been evaluated in 2 studies (Sharma et al., 2007; Lv et al., 2013). Sharma et al performed a feasibility trial with Olympus Q240Z with quadrantic examination of the distal 5cm by WLE then NBI in n=50 GORD (EO n=30; NERD n=20) and controls (n=30). Similar to Fock et al, the presence of microerosions and hypervascularity was significantly higher amongst GORD. IPCL number and morphology of tortuosity, dilatation were seen significantly more in GORD versus control. These findings were consistent in independent comparison of EO and NERD versus controls. ROC analysis thresholds for best sensitivity and specificity (respectively) for NERD were maximum ipcl/field 131 (90%, 70%), min 99 (85%, 70%) and average 117 (90%, 70%) (Sharma et al., 2007).
Lv et al used the Olympus GIF-H260Z to evaluate NERD (n=40), EO (n=40), Barrett's (n=40) and healthy controls (n=40). IPCL number, morphology (prolonged/dilated/tortuous), microerosions, round pit pattern above or below the SCJ, were recorded as features of reflux. Significant differences were found with increased IPCL number, microerosions, non-round pit patterns below the SCJ in GERD (NERD/EO and BE) patients compared to controls and fewer microerosions in NERD patients compared to RE (Lv et al., 2013).
The definition of NERD in all studies to date, however, is variable and largely based on symptom evaluation, response to PPI and the absence of mucosal lesions at endoscopic examination without standardisation using pH studies.
Artificial Intelligence
To date there is one study evaluating the use of ANNs in predicting GORD based on 45 variables including demographics, medical history, health status, symptoms scores. All patients underwent OGD, 24-pH studies performed in those with no mucosal lesion at endoscopy: 103 GORD patient (62 with reflux oesophagitis and 41 with AET>5%) and n=56 FH patients GORD. The ANN demonstrated an accuracy of 100% compared to 78% using conventional statistical regression analysis (Pace et al., 2005). While these are optimistic findings, the proportion of training and test data used was not specified and further evaluation with larger datasets is clearly warranted. There are no image-driven AI models for the diagnosis of GORD to date. Machine learning with endoscopic images is a pathway of great interest as described in section 1.7.7, with IPCLs as a potential target, based on previous studies of NBI for the diagnosis of GORD. CNNs involving IPCL detection and morphology have been recently reported in the context of a pilot study for the computer assisted diagnosis of oesophageal early squamous cell cancer using segmentation technology with accuracy matching expert endoscopists (Zhao et al., 2018). The image segmentation technique of Adaptive Local Thresholding has been demonstrated to be useful in vessel detection in retinal photographs making this as attractive technique for IPCLs (Jiang and Mojon, 2003).
Inclusion Criteria:
Exclusion Criteria: