Evaluation of recent lightweight deep learning architectures for lung cancer CT classification

While numerous large and complex deep learning architectures continue to be developed for medical imaging, clinical adoption remains limited to a small number of established models. Recent lightweight architectures, despite showing promise in computer vision tasks, are underutilized or have never been applied to medical imaging applications, particularly lung cancer classification. This study evaluates the performance of recently developed lightweight models that have received limited attention in medical imaging tasks, establishing comprehensive baseline comparisons to guide evidence-based selection for clinical deployment in resource-constrained environments.

Using CT images, we assessed three lightweight pre-trained models-MobileOne-S0, FastViT-S12, and MambaOut-Femto for lung cancer categorization. Performance measures (accuracy, AUC) and efficiency measures (inference time, number of parameters) were contrasted. Used were a public dataset (95 cases) and a private dataset (274 cases). Resampling and data augmentation constituted part of image preparation. Five-fold cross-validation helped to validate model performance.

With the lowest inference time and modest parameters, MambaOut-Femto displayed the best efficiency. While FastViT-S12 had the largest memory usage, MobileOne-S0 used fewer parameters. On Dataset 1, MambaOut-Femto obtained a mean accuracy of 0.896 ± 0.014 and an (Area under the curve) AUC of 0.972 ± 0.004; on Dataset 2, accuracy was 0.916 ± 0.040. When compared to traditional models like ResNet and Swin Transformer on the same datasets and under the same hyperparameters, the lightweight models outperformed them with significantly lower memory usage and fewer FLOPs.

The lightweight models demonstrated superior efficiency and comparable performance to traditional models, making them ideal for deployment in low-resource settings where computational resources are limited. These findings highlight the potential for practical use in clinical workflows, overcoming barriers associated with traditional models.