COMPUTER-AIDED DIAGNOSIS IMPROVES CHARACTERIZATION OF BARRETT'S NEOPLASIA BY GENERAL ENDOSCOPISTS.

BACKGROUND & AIMS: Characterization of visible abnormalities in Barrett esophagus (BE) patients can be challenging, especially for unexperienced endoscopists. This results in suboptimal diagnostic accuracy and poor inter-observer agreement. Computer-aided diagnosis (CADx) systems may assist endoscopists. We aimed to develop, validate and benchmark a CADx system for BE neoplasia. METHODS: The CADx system received pretraining with ImageNet with consecutive domain-specific pretraining with GastroNet which includes 5 million endoscopic images. It was subsequently trained and internally validated using 1,758 narrow-band imaging (NBI) images of early BE neoplasia (352 patients) and 1,838 NBI images of non-dysplastic BE (173 patients) from 8 international centers. CADx was tested prospectively on corresponding image and video test sets with 30 cases (20 patients) of BE neoplasia and 60 cases (31 patients) of non-dysplastic BE. The test set was benchmarked by 44 general endoscopists in two phases (phase 1: no CADx assistance; phase 2: with CADx assistance). Ten international BE experts provided additional benchmark performance. RESULTS: Stand-alone sensitivity and specificity of the CADx system were 100% and 98% for images and 93% and 96% for videos, respectively. CADx outperformed general endoscopists without CADx assistance in terms of sensitivity (p=0.04). Sensitivity and specificity of general endoscopist increased from 84% to 96% and 90 to 98% with CAD assistance (p<0.001), respectively. CADx assistance increased endoscopists' confidence in characterization (p<0.001). CADx performance was similar to Barrett experts. CONCLUSION: CADx assistance significantly increased characterization performance of BE neoplasia by general endoscopists to the level of expert endoscopists. The use of this CADx system may thereby improve daily Barrett surveillance.

Impact of visual features on the segmentation of gastroenterology images using normalized cuts.

Gastroenterology imaging is an essential tool to detect gastrointestinal cancer in patients. Computer-assisted diagnosis is desirable to help us improve the reliability of this detection. However, traditional computer vision methodologies, mainly segmentation, do not translate well to the specific visual characteristics of a gastroenterology imaging scenario. In this paper, we propose a novel method for the segmentation of gastroenterology images from two distinct imaging modalities and organs: chromoendoscopy (CH) and narrow-band imaging (NBI) from stomach and esophagus, respectively. We have used various visual features individually and their combinations (edgemaps, creaseness, and color) in normalized cuts image segmentation framework to segment ground truth datasets of 142 CH and 224 NBI images. Experiments show that an integration of edgemaps and creaseness in normalized cuts gives the best segmentation performance resulting in high-quality segmentations of the gastroenterology images.

Invariant Gabor texture descriptors for classification of gastroenterology images.

Automatic classification of lesions for gastroenterology imaging scenarios poses novel challenges to computer-assisted decision systems, which are mostly attributed to the dynamics of the image acquisition conditions. Such challenges demand that automatic systems are able to give robust characterizations of tissues irrespective of camera rotation, zoom, and illumination gradients when viewing the inner surface of the gastrointestinal tract. In this paper, we study the invariance properties of Gabor filters and propose a novel descriptor, the autocorrelation Gabor features (AGF). We show that our proposed AGF is invariant to scale, rotation, and illumination changes in the images. We integrate these new features in a texton framework (Texton-AGF) to classify images from two complementary gastroenterology imaging scenarios (chromoendoscopy and narrow-band imaging) broadly into three different groups: normal, precancerous, and cancerous. Results show that they compare favorably to using state-of-the-art texture descriptors for both imaging modalities.

Endoscopic assessment and grading of Barrett's esophagus using magnification endoscopy and narrow-band imaging: accuracy and interobserver agreement of different classification systems (with videos).

BACKGROUND: Three different classification systems for the evaluation of Barrett's esophagus (BE) using magnification endoscopy (ME) and narrow-band imaging (NBI) have been proposed. Until now, no comparative and external evaluation of these systems in a clinical-like situation has been performed. OBJECTIVE: To compare and validate these 3 classification systems. DESIGN: Prospective validation study. SETTING: Tertiary-care referral center. Nine endoscopists with different levels of expertise from Europe and Japan participated as assessors. PATIENTS: Thirty-two patients with long-segment BE. INTERVENTIONS: From a group of 209 standardized prospective recordings collected on BE by using ME combined with NBI, 84 high-quality videos were randomly selected for evaluation. Histologically, 28 were classified as gastric type mucosa, 29 as specialized intestinal metaplasia (SIM), and 27 as SIM with dysplasia/cancer. Assessors were blinded to underlying histology and scored each video according to the respective classification system. Before evaluation, an educational set concerning each classification system was carefully studied. At each assessment, the same 84 videos were displayed, but in different and random order. MAIN OUTCOME MEASUREMENTS: Accuracy for detection of nondysplastic and dysplastic SIM. Interobserver agreement related to each classification. RESULTS: The median time for video evaluation was 25 seconds (interquartile range 20-39 seconds) and was longer with the Amsterdam classification (P < .001). In 65% to 69% of the videos, assessors described certainty about the histology prediction. The global accuracy was 46% and 47% using the Nottingham and Kansas classifications, respectively, and 51% with the Amsterdam classification. The accuracy for nondysplastic SIM identification ranged between 57% (Kansas and Nottingham) and 63% (Amsterdam). Accuracy for dysplastic tissue was 75%, irrespective of the classification system and assessor expertise level. Interobserver agreement ranged from fair (Nottingham, κ = 0.34) to moderate (Amsterdam and Kansas, κ = 0.47 and 0.44, respectively). LIMITATION: No per-patient analysis. CONCLUSIONS: All of the available classification systems could be used in a clinical-like environment, but with inadequate interobserver agreement. All classification systems based on combined ME and NBI, revealed substantial limitations in predicting nondysplastic and dysplastic BE when assessed externally. This technique cannot, as yet, replace random biopsies for histopathological analysis.