ACLA-Net: attention-guided cross-level alignment network for MRI-based prostate cancer lesion segmentation

Benefiting from its radiation-free property and excellent soft-tissue contrast, magnetic resonance imaging (MRI) has become a crucial modality for prostate cancer diagnosis. However, automatic segmentation of prostate cancer lesions in MRI is still a challenging task due to the considerable variations in lesion morphology and scale, as well as the often indistinct boundaries between tumors and surrounding tissues. To address these challenges, we propose ACLA-Net, an attention-guided cross-level alignment network tailored for MRI-based prostate cancer lesion segmentation.

Specifically, ACLA-Net is built upon the U-Net architecture. To enhance the interaction and alignment of encoder features across different semantic levels, a cross-level feature alignment module (CLFAM) is introduced into the second, third, and fourth skip connections. Meanwhile, a spatial attention-guided fusion module (SAFM) is incorporated into the first skip connection to emphasize discriminative spatial cues and improve the fusion of shallow features. At the bottleneck stage, a SE-guided context aggregation module (SECAM) is employed to capture richer global contextual information and strengthen high-level semantic representation. Furthermore, the conventional double convolution blocks in the encoder are replaced with the SE-guided multi-kernel depthwise module (SEMKDM), which improves multi-scale feature extraction while reducing computational redundancy.

To evaluate the effectiveness and generalization ability of the proposed method, extensive experiments were conducted on two self-constructed prostate cancer MRI datasets, namely ProstateCancer-T2WI and ProstateCancer-ADC, as well as the public DDTI thyroid dataset. On ProstateCancer-T2WI, ACLA-Net achieved Dice, MCC, and Jaccard scores of 0.7798, 0.7782, and 0.6403, while on ProstateCancer-ADC, it obtained corresponding scores of 0.8332, 0.8328, and 0.7159. Furthermore, on the public DDTI dataset, ACLA-Net attained Dice, MCC, and Jaccard scores of 0.7752, 0.7438, and 0.6338. In addition, ablation studies were carried out to verify the effectiveness of each proposed module. The experimental results demonstrate that ACLA-Net achieves reliable segmentation performance on prostate cancer MRI and favorable generalization on the public dataset, suggesting its potential to provide useful support for lesion delineation and subsequent clinical assessment.