Akademik Veri Yönetim Sistemi
16. Ulusal ve 3. Uluslararasi Endoskopik Laparoskopik-Robotik Cerrahi Kongresi ve 21. MMESA Kongresi, Antalya, Türkiye, 26 - 29 Ekim 2023, ss.1
Introduction:
An essential purpose of upper gastrointestinal system endoscopy is detecting Helicobacter pylori(H.pylori) infection through visual examination of the gastric mucosa, which plays significant role in the development of gastric cancer. However, there are currently no proven techniques for optically diagnosing H. pylori infection utilizing endoscopic images. Therefore, endoscopic biopsy is necessary for definitive diagnosis. Artificial intelligence is increasingly used for image classification and object detection. Our goal was to create a convolutional neural network(CNN). This deep learning method can identify specific features in endoscopic images.
Methods:
248 patients with gastric symptoms underwent endoscopic examination, provided written consent, and were retrospectively recruited. H.pylori infection was assessed from a biopsy taken from the antrum. Endoscopic RGB image inputs were resized to 224x224x3. Min-max normalization was utilized for preprocessing. The naive random over-sampling method was used to overcome the class imbalance problem. A hyperparameter optimization framework was used to choose the optimal architecture for dataset. The model consisted of one-block 2D CNN, flatten, dropout and dense layers. H.pylori infection status was obtained from the last dense layer of the sigmoid activation function. Binary cross-entropy loss function and accuracy were used in the validation cohort to evaluate model performance.
Results:
One hundred fourteen patients with 114 endoscopic images from the antrum were in the H.pylori(+) group, while 134 patients with 144 images were in the H.pylori(-) group. Among 49 images in the validation cohort, CNN diagnosed 18 images as positive and 31 as negative. The sensitivity, specificity, and accuracy of the CNN for the detection of H.pylori were 73.6 %, 86.6 %, and 81.6 %, respectively.
Conclusion: Deep learning algorithms applied to endoscopic images might help reveal underlying H.pylori infection, which can offer potential guidance in treatment decision. Future studies will validate deep learning results in a larger cohort.