Segmentation of lung cancer-caused metastatic lesions in bone scan images using self-defined model with deep supervision. (January 2023)

• Record Type: Journal Article
• Authors: Cao, Yongchun; Liu, Liangxia; Chen, Xiaoyan; Man, Zhengxing; Lin, Qiang; Zeng, Xianwu; Huang, Xiaodi
• Is Part Of: Biomedical signal processing and control. Volume 79(2023)Part 1
• Publication Date: 2023-01
• Subjects: Bone scan -- Skeletal metastasis -- Lung cancer -- Image segmentation -- Convolutional neural network; SPECT Single Photon Emission Computed Tomography -- PET Positron Emission Tomography -- 99mTc-MDP 99mTc-methylene diphosphonate -- CNN Convolutional Neural Network -- CPA Class Pixel Accuracy -- DSC Dice Similarity Coefficient; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2022.104068
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: Automated segmentation of lesions metastasized from lung cancer in thorax is first studied with SPECT bone scan images. A view aggregation method is developed to enhance the areas within a SPECT image that may denote the metastatic lesions. A self-defined end-to-end segmentation network is proposed by introducing deep supervision into the encoder-decoder network to automatically identify and delineate metastatic lesions. Extensive experiments conducted on clinical SPECT bone scan images show the proposed model's feasibility and superiority as compared to the classical image segmentation models. Abstract: To automatically identify and delineate metastatic lesions in low-resolution bone scan images, we propose a deep learning-based segmentation method in this paper. In particular, the view aggregation in this method uses a pixel-wise addition to enhance the regions with high uptake of the radiopharmaceutical. The operation of view aggregation augments images for the lesion segmentation task. By following the structure of the encoder-decoder with deep supervision, our model is an end-to-end segmentation network that consists of two sub-networks of feature extraction and pixel classification. As such, the hieratical features of bone scan images can be learned by the feature extraction sub-network. The pixels in metastasis areas within a feature map are then identified and delineated by the pixel classification sub-network. The results of experiments on clinical bone scan images show that the proposed model performs well in segmenting metastatic lesions automatically, obtaining a mean score of 0.6556 on DSC (Dice Similarity Coefficient). However, more bone scan images enable our model to learn better representative features of metastatic lesions, for further improving the performance of deep learning-based lesion segmentation.