IDDF2023-ABS-0229 U-net-based deep learning model for accurate helicobacter pylori delineation in gastric biopsy specimens

Background Helicobacter pylori (H. pylori) is a gram-negative spiral bacterium whose presence is a significant risk factor contributing to gastric cancer development. However, accurately identifying H. pylori in biopsy specimens can be challenging for pathologists due to the bacterium’s small size, which can lead to inconsistencies and misdiagnosis. To address this issue, we propose to develop a deep-learning model for the H. pylori-infected region delineation in gastric biopsy specimens. Our goal was to support pathologists in identifying and diagnosing H. pylori infection, thereby improving the accuracy of diagnosis for patients with gastric cancer.MethodsWe collected haematoxylin and eosin (H&E) stained slides of five gastric biopsy cases positive for H. pylori and digitized them using our Hamamatsu NanoZoomer S210 whole slide imaging system at 40x magnification (0.23 µm/pixel). The areas positive for H. pylori on the whole slide images (WSI) were annotated, and 512 x 512-pixel image tiles without overlap were extracted for model development, resulting in a total of 1,717 image tiles containing gastric tissue. We trained a modified U-Net with ResNet34 backbone and Lovász-Softmax loss function. The proposed U-Net was configured with a batch size of 128, 400 epochs, Adam optimizer, and a 1.00-04 learning rate. We evaluated our model based on the Intersection of Union (IoU) and Sørensen–Dice coefficient (DICE).ResultsOur U-Net-based model demonstrated excellent performance for H. pylori infection region delineation, with an IoU of 0.7805 and DICE of 0.8767. Figure 1 illustrates that the model significantly highlights most of the H. pylori-infected regions. (IDDF2023-ABS-0229 Figure 1. Representative original tiled H&E-stained whole slide images (WSI) with H. pylori infection (A – B). The predicted (highlighted in yellow) and ground truth (highlighted in blue) regions generated from our proposed U-Net model are shown (C – D). The black and red arrows indicate the H. pylori present within the ground truth and predicted regions from our proposed U-Net model, respectively)Abstract IDDF2023-ABS-0229 Figure 1Representative original tiled H&E-stained whole slide images (WSI) with H. pylori infection (A – B). The predicted (highlighted in yellow) and ground truth (highlighted in blue) regions generated from our proposed U-Net model are shown (C – D). The black and red arrows indicate the H. pylori present within the ground truth and predicted regions from our prop