Deep learning-based classification of lung adenocarcinoma subtypes in histopathological images using DS-EffNet

Area of Science:

• Oncology
• Medical Imaging
• Computer Science

Background:

• Lung adenocarcinoma (LADC) subtype classification is crucial for understanding disease heterogeneity and guiding treatment.
• Histopathological image analysis presents challenges due to complexity and heterogeneity.

Purpose of the Study:

• To develop an efficient deep learning model for accurate LADC subtype classification.
• To enhance feature extraction and modeling of complex pathological patterns in histopathological images.

Main Methods:

• Integration of Depthwise Separable Residual Block (DSResBlock), RefConv, Channel Attention Pooling (CAP), and Multidimensional Collaborative Attention (MCA) modules into EfficientNetV2-S.
• Development of the DS-EffNet model for optimized feature extraction and pattern modeling.

Main Results:

• The DS-EffNet model achieved 95.1% accuracy, 0.938 F1-score, and 0.994 AUC on the primary dataset.
• Achieved 100% generalization accuracy on the LC25000 dataset, demonstrating cross-institutional performance.
• Ablation studies confirmed the synergistic contribution of each module, especially MCA for complex features and RefConv for computational efficiency.

Conclusions:

• The proposed DS-EffNet model offers a novel and efficient approach for LADC subtype classification.
• The model's performance suggests potential as a tool to assist pathologists in rapid subtype information delivery.
• This study provides a new design paradigm for medical image classification applicable to other histological tasks and aids in treatment stratification.