MEAA-Net: Memory-Efficient Asymmetric Attention for Resource-Constrained Lung Nodule Classification

Early detection and classification of lung nodules is critical for improving lung cancer prognosis, yet existing deep learning models often exceed the memory budgets of typical clinical deployment hardware, hindering real-time deployment. We introduce MEAA-Net, a memory-efficient asymmetric attention network that processes intermediate feature maps into batch-wise partitions and leverages gradient checkpointing to reduce peak training memory requirements. Five architectural variants (ranging from 61 K to 14.65 million parameters and 0.76 to 3.73 GFLOPs) were trained under a unified protocol and evaluated on the validation splits of the LUNA16 and LIDC-IDRI public datasets. Our smallest model (61 K parameters; 0.76 GFLOPs) achieved 80.8% accuracy on LUNA16 and 65.4% on LIDC-IDRI, corresponding to about 12 accuracy points per million parameters and delivering single-slice inference in 11 ms on a standard CPU. Compared with ResNet-18 (11.24 M parameters), this variant reduces parameter count by 184 × while keeping mean validation accuracy within 2 percentage points under the same evaluation protocol. MEAA-Net’s highly compact design and competitive performance enable real-time, hardware-friendly lung nodule classification in resource-constrained clinical environments, paving the way for broader adoption of AI-assisted diagnostics in primary and community care settings.