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

The classification of lung adenocarcinoma (LADC) subtypes is important for understanding disease heterogeneity and has potential implications for diagnosis and treatment planning, yet the complexity and heterogeneity of histopathological images pose significant challenges for automated classification. This study proposes an efficient deep learning model that integrates Depthwise Separable Residual Block (DSResBlock), RefConv, Channel Attention Pooling (CAP), and Multidimensional Collaborative Attention (MCA) modules into the EfficientNetV2-S architecture to improve subtype classification of LADC histopathological images. DS-EffNet model optimizes feature extraction and modeling of complex pathological patterns through depthwise separable convolutions and attention mechanisms. Experimental results demonstrate that the model achieves an accuracy of 95.1 %, an F1-score of 0.938, and an area under the curve (AUC) of 0.994 on the primary experimental dataset, outperforming other baseline models. Furthermore, the model attains 100 % generalization accuracy on the LC25000 dataset, indicating promising cross-institutional performance. Ablation studies demonstrate the synergistic contributions of each module, with MCA particularly effective in modeling complex features and RefConv significantly reducing computational complexity. This study provides a novel design paradigm for medical image classification, extendable to other histological tasks, and may assist pathologists by providing rapid subtype-level information, potentially supporting future diagnostic research and aiding treatment stratification studies.