Beeres, M.H (2020) The development of a deep learning computeraided detection system to detect Barrett neoplasia using videos. thesis, Medicine.
Full text available on request.Abstract
Background: Barrett esophagus (BE) is a premalignant condition predisposing to esophageal adenocarcinoma (EAC). If EAC is diagnosed at an early stage it can be removed endoscopically, leading to an excellent prognosis. However, early BE neoplasia is difficult to detect and therefore, often missed during endoscopic surveillance. A computer-aided detection (CAD) system could be beneficial to assist the endoscopist to recognize early neoplasia. The aim of our study was to develop a video-based CAD system for endoscopic detection of BE neoplasia. Methods: The CAD system was tested with images and videos separated in four datasets. The first two datasets were developed and used in previous studies. The third dataset contained 589 images of BE neoplasia and 378 with non-dysplastic BE (NDBE) in 213 patients and contained a wide variety in image quality. Dataset four contained 21 videos of BE neoplasia and 22 NDBE videos in 43 patients. All images and videos had corresponding histopathology which was used as golden standard. The primary outcome measures were the diagnostic performance of the CAD system per-image/per video-frame and per-patient. Defined by the accuracy, sensitivity and specificity of BE neoplasia detection. In addition, the diagnostic performance was also calculated for CAD system predictions with a high level of confidence and assessment time. In dataset 4 the delineation of the CAD system was compared with the delineation of the expert resulting in segmentation performance. Results: In dataset 3 accuracy, sensitivity and specificity of the CAD system per-patient were 69% (95% CI 63-75), 59% (95% CI 51-67) and 89% (95% CI 81-96). The diagnostic performance on dataset 4 per-patient was 95% (95% CI 89-100), 91% (95% CI 77-100) and 100% (95% CI 100), respectively. In both datasets, the diagnostic performance improved considerable in a range of 5-10% in high confidence level. The CAD system delineated 96% (204/212) video-frames correctly. The assessment time for the CAD system to classify and delineate a video-frame was 0.26 seconds (SD 0.021) Conclusions: This was one of the first studies where videos were used to evaluate CAD system performance for BE neoplasia detection. The results on images (dataset 3) were lower than expected, most likely due to low image quality. On videos (dataset 4), the CAD system detected neoplasia with an excellent diagnostic performance and fast corresponding assessment time. To improve the generalizability of the CAD system, more data acquisition and prospective trials in daily practice are needed.
| Item Type: | Thesis (UNSPECIFIED) |
|---|---|
| Supervisor name: | Nagengast, W.B and Struyvenberg, M.R |
| Faculty: | Medical Sciences |
| Date Deposited: | 15 Aug 2023 11:23 |
| Last Modified: | 15 Aug 2023 11:23 |
| URI: | https://umcg.studenttheses.ub.rug.nl/id/eprint/3626 |
Actions (login required)
 |
View Item |