MECF-Net: A prostate cancer lymph node metastasis classification method based on 18F-PSMA-1007 and 18F-FDG dual-tracer PET/CT image feature optimization

Abstract

Prostate cancer is one of the most common cancers in men, and the presence of lymph node metastasis is critical for determining treatment options and assessing prognosis. ¹⁸F-FDG is a widely used tracer in PET imaging for tumor diagnosis, while ¹⁸F-PSMA-1007 typically exhibits high specificity in prostate cancer cells. Therefore, we propose a novel model for classifying prostate cancer lymph node metastasis, MECF-Net, which integrates ¹⁸F-PSMA-1007 and ¹⁸F-FDG PET/CT images. Specifically, to enhance feature perception at different channel levels and multi-scale spatial dimensions, we propose the Multi-Scale Feature Extraction (MSFE) branch, which combines Squeeze-and-Excitation Attention with a newly designed Multi-Scale Spatial Enhanced Attention (MSEA). The MSEA extracts spatial feature information of tumors at various scales by employing global average pooling and max pooling aggregation operations at multiple scales. Furthermore, we introduce a Local-Global Feature Complementary Fusion (LGFCF) branch, which constructs a series of complementary fusion blocks as basic units. These blocks consist of concatenated multi-scale grouped convolutions and point-wise convolutions, enabling the complementary extraction of intra-group local spatial features and inter-channel global features. Finally, at the end of MECF-Net, we design a Multi-Feature Adaptive Fusion (MF-AF) module, based on a dynamic weight allocation mechanism, to fuse the features extracted from different branch sub-networks. Our experimental results on a private dual-tracer ¹⁸F-PSMA-1007/¹⁸F-FDG PET/CT dataset and a public single-tracer ¹⁸F-FDG PET/CT dataset demonstrate the effectiveness of MECF-Net, which achieved 0.9269 and 0.885 in ACC, respectively, 0.9156 and 0.8838 in AUC, respectively, demonstrating superior performance compared to state-of-the-art networks, as well as generalizability on the single-tracer dataset.