TLDR: CoAtNeXt is a new hybrid deep learning model combining convolutional and transformer architectures, specifically designed for gastric tissue image classification. It enhances the CoAtNet architecture by integrating attention-enhanced ConvNeXtV2 blocks. Evaluated on two public datasets (HMU-GC-HE-30K for eight-class and GasHisSDB for binary classification), CoAtNeXt achieved superior accuracy (96.47% and 98.29% respectively) and AUC scores compared to existing CNN and Vision Transformer models, while also being computationally efficient. This model shows significant potential to improve diagnostic accuracy and reduce pathologist workload in gastric disease detection.
Early and accurate diagnosis of gastric diseases is paramount in preventing severe health outcomes. While traditional histopathological examination remains the gold standard, it is a labor-intensive process, prone to variations among pathologists, and can sometimes miss critical findings due to its manual nature. These limitations underscore a pressing need for automated, reliable, and efficient methods for analyzing gastric tissue.
Artificial intelligence (AI) has emerged as a powerful tool in medical image analysis, with Convolutional Neural Networks (CNNs) excelling at local feature extraction and Vision Transformers (ViTs) demonstrating prowess in capturing global contextual relationships. However, many existing AI approaches for gastric tissue analysis often struggle to effectively combine both local and global feature modeling simultaneously. CNNs might miss broader contextual information, while Transformers can overlook fine local details. Furthermore, many studies have focused primarily on binary classification (e.g., normal vs. abnormal), limiting the models’ ability to generalize across multiple gastric tissue types.
To address these challenges, a novel hybrid model named CoAtNeXt has been developed for the classification of gastric tissue images. This innovative model builds upon the established CoAtNet architecture, which itself is a blend of convolutional and transformer mechanisms. The key enhancement in CoAtNeXt involves replacing CoAtNet’s standard MBConv layers with advanced ConvNeXtV2 blocks. These ConvNeXtV2 blocks are further augmented with a Convolutional Block Attention Module (CBAM), which significantly improves local feature extraction by allowing the model to adaptively focus on important regions in both channel and spatial dimensions.
The CoAtNeXt architecture is meticulously scaled to strike a balance between computational efficiency and classification performance. In its early stages, the model leverages the enhanced ConvNeXtV2 blocks to learn robust local features. In later stages, it retains the powerful relative positional transformer blocks from CoAtNet to effectively model long-range contextual relationships. This hybrid design ensures that CoAtNeXt can capture both the intricate fine-grained details of histopathological images and the broader structural patterns.
The model’s effectiveness was rigorously evaluated on two widely recognized public datasets: HMU-GC-HE-30K and GasHisSDB. The HMU-GC-HE-30K dataset is used for a more complex eight-class classification task, distinguishing between various gastric tissue types such as adipose tissue, necrotic debris, lymphocyte accumulation, mucus, muscle tissue, normal mucosa, connective tissue, and tumor tissue. The GasHisSDB dataset, on the other hand, is utilized for binary classification, differentiating between normal and abnormal gastric tissues.
The results were highly promising. On the HMU-GC-HE-30K dataset, CoAtNeXt achieved an impressive 96.47% accuracy, 96.60% precision, 96.47% recall, 96.45% F1 score, and 99.89% AUC. For the binary classification on GasHisSDB, it reached 98.29% accuracy, 98.07% precision, 98.41% recall, 98.23% F1 score, and 99.90% AUC. These figures demonstrate that CoAtNeXt consistently outperformed all ten CNN and ten ViT models tested in the study, as well as surpassing previous research findings in the literature.
Beyond its superior performance, CoAtNeXt also boasts a highly efficient structure. With only 18.8 million parameters and a compact size of 110 MB, it is significantly lighter than many other high-performing models like SwinTransformerV2 (49.7M parameters / 222 MB) and PVTv2 (62.6M parameters / 241 MB). This computational efficiency, combined with its high accuracy, makes CoAtNeXt a practical and valuable tool for potential clinical applications, where resource constraints and real-time processing are often critical considerations.
Also Read:
- Mobile Phone Images Power AI for Early Arsenicosis Detection
- Enhancing Brain Tumor Segmentation with EMCAD: A Focus on Efficiency and Multi-scale Attention
In conclusion, CoAtNeXt represents a significant advancement in the field of AI-assisted gastric tissue classification. Its ability to effectively integrate local and global feature extraction, coupled with its robust performance across both binary and multi-class tasks, highlights its potential to enhance diagnostic accuracy, reduce the workload on pathologists, and ultimately contribute to earlier and more effective treatment of gastric diseases. For more details, you can refer to the full research paper: CoAtNeXt: An Attention-Enhanced ConvNeXtV2–Transformer Hybrid Model for Gastric Tissue Classification.
