Bone tumor necrosis rate detection in few-shot X-rays based on deep learning. (December 2022)

• Record Type: Journal Article
• Authors: Xu, Zhiyuan; Niu, Kai; Tang, Shun; Song, Tianqi; Rong, Yue; Guo, Wei; He, Zhiqiang
• Is Part Of: Computerized medical imaging and graphics. Volume 102(2022)
• Publication Date: 2022-12
• Subjects: Bone tumor necrosis -- Deep learning -- Few-shot samples -- Generative adversarial network -- Time series; Diagnostic imaging -- Periodicals; Imaging systems in medicine -- Periodicals; Diagnosis, Radioscopic -- Data processing -- Periodicals; Diagnostic Imaging -- Periodicals; Imagerie pour le diagnostic -- Périodiques; Diagnostic imaging; Periodicals; Electronic journals; Electronic journals; 616.0754
• DOI: 10.1016/j.compmedimag.2022.102141
• Languages: English
• ISSNs: 0895-6111

Abstract

Abstract: Although biopsy-based necrosis rate is a golden standard for reflecting the sensitivity of bone tumor and guiding postoperative chemotherapy, it requires biopsy which is invasive and time-consuming. In this paper, we develop a new necrosis rate detection method using time series X-ray images instead of biopsy. To overcome the limitations of few-shot samples, the proposed method utilizes a Generative Adversarial Network with Long Short-term Memory to generate time series X-ray images. For further data expansion, an image-to-image translation network is applied for producing the initial images. These augmented data are treated as the training set of a 3D-Convolutional Neural Network classification model. Our method expands the few-shot bone tumor X-rays by 10 times, and approaches the necrotic rate classification result of biopsy, which is the state-of-the-art technique in the detection of few-shot bone tumor necrosis rate. Furthermore, it provides an efficient method to investigate the bone tumor necrosis rate in few-shot samples. Highlights: A deep learning method is proposed to detect necrosis rate with few-shot X-rays. The Conv-LSTM-GAN network in our method is designed to generate bone tumor images. Our method expands bone tumor images by 10 times and achieves biopsy-like results.