BSci-Seg: A bone scintigram segmentation model for accurately detecting and delineating metastasis lesions using convolutional neural networks

Metastatic lesion segmentation is a crucial task for diagnosis and follow-up assessments of patients with malignancies. Recent advances in Convolutional Neural Networks have introduced promising approaches for bone metastasis lesion segmentation. While previous efforts have focused on enhancing the performance of U-Net-based models, challenges persist regarding clinical interpretability and lesion sensitivity. In this paper, we propose a novel bone scintigraphy segmentation model, BSci-Seg, designed to address these challenges by leveraging domain-specific patterns and improving both performance and interpretability. BSci-Seg is built upon a classical encoder-decoder architecture and incorporates a paired dual-sampling scheme (PDSS), multiple receptive-field attention (MRFA), and a customized loss function. The network employs PDSS to relevant layers during both encoding and decoding, and utilizes MRFA modules in the feature encoding stage to enhance differential representation across regions. Experimental evaluations on 286 SPECT bone scintigrams show significant improvements, with a 4.53 % increase in Dice Similarity Coefficient (DSC) and a 9.90 % increase in Recall. Additionally, comparisons with existing models for bone metastasis lesion segmentation demonstrate the superior performance of BSci-Seg. Comprehensive ablation studies and detailed case analyses further validate the effectiveness of the model, laying the groundwork for further research.