JMIR Publications

ABSTRACT

Background:

Cytomegalovirus (CMV) reactivation is common among patients with severe ulcerative colitis (UC), resulting in poorer prognoses than patients without CMV reactivation. The principal diagnostic approach for CMV involves biopsies, which are time-consuming and present challenges for early detection. To address this issue, our study utilizes deep learning to differentiate CMV from severe UC using endoscopic imaging, thereby enabling early CMV diagnosis. Materials and

Methods:

In this study, we examined 86 endoscopic images employing an ensemble of deep learning models, notably Densenet 121 pre-trained on ImageNet, to discriminate between cases of UC with and without CMV complications. Extensive preprocessing and test-time augmentation (TTA) techniques were applied to enhance the effectiveness of the models. Evaluation of the models' performance included metrics such as accuracy, precision, recall, F1 score, receiver operating characteristic (ROC) curves, and AUC values, highlighting the potential of deep learning to improve non-invasive gastroenterology diagnostics.

Results:

An ensemble of four models augmented with TTA demonstrated superior performance in classifying UC endoscopic images. It attained an accuracy of 0.83, precision of 0.83, recall of 0.91, and an F1-score of 0.87. These metrics underscore the ensemble's reliability and well-rounded performance. Particularly noteworthy is the substantial decline in performance metrics observed in models without TTA, highlighting the critical role of TTA.

Conclusions:

Our findings underscore the effectiveness of deep learning models in distinguishing CMV from severe UC in endoscopy images, providing a viable approach for non-invasive diagnostics and timely therapeutic interventions