Investigating Convolution-Attention Model for Bone Scan Image Segmentation

Bone scan image segmentation is a crucial step in the early detection of a tumor spreading across the human body. By dividing each bone region, the Bone Scan Index can be analyzed for further follow-up against cancer. Much research has been done in this field, including the uses of Computer Aided Diagnosis (CAD) in an application developed by EXINI, the development of Active Shape Model (ASM), and Constrained Local Model (CLM) as a further development of ASM. However, these models still rely on landmark points for their training phase instead of using masks for the annotations. A recent convolution model, DeepLabv3+, relies on a convolution mechanism to extract local features. A new approach using a pure transformer, Segmenter, can extract global features in a parallel process. A combination of convolution and attention, Dual Attention Network (DANet) uses a similar backbone from DeepLabv3+ and implements attention modules to capture long-range contextual information from the input image. In this paper, a bone scan image segmentation system using DANet will be proposed. All models are trained using the bone scan dataset divided into anterior and posterior groups. The annotation is composed of 12 different classes of bone regions. The results show that the convolution-attention approach of DANet outperformed existing models in both the anterior and posterior sections. A performance of 76.85% mIoU is achieved in the anterior section, and 80.99% mIoU is achieved in the posterior section.