Improving Classification Performance in Gastric Disease through Realistic Data Augmentation Technique Based on Poisson Blending

초록

Gastric cancer ranks fifth in incidence and fourth in mortality worldwide; thus, regular gastroscopies are necessary to reduce its incidence and mortality rates. However, diagnosing early-stage gastric cancer is challenging owing to subtle changes in the mucosa. Therefore, this study proposes a deep learning-based computer-aided diagnosis system (CADx). CADx, trained using EfficientNetV2, can distinguish early-stage gastric cancer as abnormal for its diagnosis, thereby providing additional opinions to specialists and reducing misdiagnosis rates. The quality and quantity of the training data influence the performance of CADx. However, collecting medical data is challenging as it requires obtaining patient and Institutional Review Board approval. Therefore, this study proposes a data-augmentation technique that synthesizes abnormal and normal endoscopic images using Poisson blending. The proposed method is compared with the previously used cut-and-paste method; evidently, the proposed method produces realistic images. The model trained using the proposed method exhibits a sensitivity and F1-score of 0.917 and 0.948 for classifying abnormal and normal cases and 0.823 and 0.823 for classifying early gastric cancer, abnormal, and normal cases, respectively. The proposed method exhibits approximately 30% improvement in performance compared with the models trained using the original dataset, thus demonstrating its effectiveness in endoscopic classification.

키워드