Interdisciplinary Sciences: Computational Life Sciences Медицинская информатика
CoAtNeXt: An Attention-Enhanced ConvNeXtV2–Transformer Hybrid Model for Gastric Tissue Classification
Аннотация
Early diagnosis of gastric diseases is essential to prevent adverse outcomes. The current gold standard, histopathological examination, is labor-intensive and subject to inter-pathologist variability, which can lead to critical findings being missed. This underscores the need for automated, reliable methods for analyzing gastric tissue. To address these challenges, CoAtNeXt, a hybrid model for tissue image classification, is introduced. The CoAtNeXt architecture, based on CoAtNet, replaces MBConv layers with enhanced ConvNeXt-V2 blocks and integrates the Convolutional Block Attention Module (CBAM) to improve local feature extraction through channel and spatial attention. This design aims to balance computational efficiency with classification performance. Evaluation on two public datasets—HMU-GC-HE-30 K for an 8-class task and GasHisSDB for binary classification—shows that CoAtNeXt achieves 96.47% accuracy, 96.60% precision, and an AUC of 0.9989 on HMU-GC-HE-30 K, and 98.29% accuracy, 98.07% precision, and an AUC of 0.9990 on GasHisSDB. CoAtNeXt outperforms the 10 CNN and 10 ViT models evaluated, indicating its potential to improve robustness and diagnostic accuracy in histopathological classification while reducing pathologists’ workload. Graphical Abstract