Integrating Non-Square Filter and Boundary Enhancement Into Encoder–Decoder Network for Lesion-Aware Segmentation of Large-Size Low-Resolution Bone Scintigrams

Background: Accurate identification of bone metastases in lung cancer is essential for effective diagnosis and treatment. However, existing methods for detecting bone metastases face significant limitations, particularly in whole-body bone scans, due to low resolution, blurred boundaries, and the variability in lesion shapes and sizes, which challenge traditional convolutional neural networks. Purpose: To accurately isolate the metastasized lesions from whole-body bone scans, we propose a lesion-aware segmentation model using deep learning techniques. Methods: The proposed model integrates lesion boundaryguided strategies, multi-scale learning, and image shape guidance into an encoder-decoder architecture network. This approach significantly improves segmentation performance in low-resolution and blurred boundary conditions while effectively managing lesion shape variability and mitigating interference from the rectangular format of the images. Results: Experimental evaluations conducted on clinical data of 274 whole-body bone scans demonstrate that the proposed model achieves a 7.45% improvement in the Dice Similarity Coefficient and a 11.75% improvement in Recall compared to specialized segmentation models for whole-body bone scans, achieving significant improvements and balanced performance across key metrics. Conclusions: This model offers a more accurate and efficient solution for identifying bone metastases in lung cancer, alleviating the challenges of deep learning-based automated analysis of low-resolution, large-size medical images of whole-body bone scans. The code is available at https://github.com/carorange/segmentation.