Segmentation of Metastasized Lesions in Bone Scintigrams Using U-Net++ with Attention Gate

Nuclear medicine imaging can effectively detect metastatic cancer, and early detection of bone metastases enables effective treatment of patients. Automatic recognition of bone scans will effectively enhance the diagnostic efficiency of physicians. To reliably identify bone metastases in bone scans, in this work, we propose an automated segmentation model based on $\text{U-Net}{++}$ model. Specifically, the proposed model filters the feature maps of skip connections propagation by integrating AG (Attention gate), which enables the model to improve the learning of features in important regions. The feature fusion method is also used to make the feature maps to send to AG that contains richer information and further enhance the learning ability of the model for bone metastases lesions. An experimental evaluation using clinical data of 306 SPECT images shows that the proposed model performs well on identifying and segmenting bone metastasis lesions in bone SPECT images, obtaining scores of 0.6221, 0.6612 and 0.5878 for DSC (Dice Similarity Coefficient), CPA (Class Pixel Accuracy) and Recall. Comparison analysis with existing classical segmentation models also demonstrates the superiority of the proposed model with various loss functions. By integrating AGs into deep segmentation network and using the feature fusion, the overall performance can be improved.

I. Introduction

Bone metastases are common in patients with solid cancers, second only to metastases from lung and liver sites in terms of incidence [1], [2] and may lead to a range of skeletal-related events (SREs) such as fractures, debilitating pain, hypercalcemia and spinal cord compression [3], [4]. These SREs are the dominant cause of morbidity and mortality affecting patients [5], [6]. From a clinical perspective, early detection of bone metastases can effectively assist physicians in treating patients [3]. Nuclear medicine imaging is an effective screening modality to evaluate bone metastases. Unlike Computed Tomography (CT), Magnetic Resonance Imaging (MRI), and Ultrasound imaging, nuclear medicine imaging can use imaging agents that enter the body to image human organs, tissue structures, and functional metabolic conditions [7] and is widely used for pathological examinations in neurology, oncology and cardiology [8]–[11].