Xi-Liang Zhu1,2, Hong-Bin Shen3, Haitao Sun4, Li-Xia Duan5, Ying-Ying Xu6,7. 1. Guangdong Provincial Key Laboratory of Medical Image Processing, School of Biomedical Engineering, Southern Medical University, Guangzhou, 510515, China. 2. Guangdong Province Engineering Laboratory for Medical Imaging and Diagnostic Technology, Southern Medical University, Guangzhou, 510515, China. 3. Institute of Image Processing and Pattern Recognition, Key Laboratory of System Control and Information Processing, Ministry of Education of China, Shanghai Jiao Tong University, Shanghai, 200240, China. 4. Neurosurgery Center, The National Key Clinical Specialty, The Engineering Technology Research Center of Education Ministry of China on Diagnosis and Treatment of Cerebrovascular Disease, Guangdong Provincial Key Laboratory on Brain Function Repair and Regeneration, The Neurosurgery Institute of Guangdong Province Zhujiang Hospital, Southern Medical University, Guangzhou, 510282, China. 5. Guangzhou Red Cross Hospital, Medical College, Jinan University, Guangzhou, 510220, Guangdong, China. dlix332000@163.com. 6. Guangdong Provincial Key Laboratory of Medical Image Processing, School of Biomedical Engineering, Southern Medical University, Guangzhou, 510515, China. yyxu@smu.edu.cn. 7. Guangdong Province Engineering Laboratory for Medical Imaging and Diagnostic Technology, Southern Medical University, Guangzhou, 510515, China. yyxu@smu.edu.cn.
Abstract
PURPOSE: Computed tomography (CT) images can display internal organs of patients and are particularly suitable for preoperative surgical diagnoses. The increasing demands for computer-aided systems in recent years have facilitated the development of many automated algorithms, especially deep convolutional neural networks, to segment organs and tumors or identify diseases from CT images. However, performances of some systems are highly affected by the amount of training data, while the sizes of medical image data sets, especially three-dimensional (3D) data sets, are usually small. This condition limits the application of deep learning. METHODS: In this study, given a practical clinical data set that has 3D CT images of 20 patients with renal carcinoma, we designed a pipeline employing transfer learning to alleviate the detrimental effect of the small sample size. A dual-channel fine segmentation network (FS-Net) was constructed to segment kidney and tumor regions, with 210 publicly available 3D images from a competition employed during the training phase. We also built discriminative classifiers to classify the benign and malignant tumors based on the segmented regions, where both handcrafted and deep features were tested. RESULTS: Our experimental results showed that the Dice values of segmented kidney and tumor regions were 0.9662 and 0.7685, respectively, which were better than those of state-of-the-art methods. The classification model using radiomics features can classify most of the tumors correctly. CONCLUSIONS: The designed FS-Net was demonstrated to be more effective than simply fine-tuning on the practical small size data set given that the model can borrow knowledge from large auxiliary data without diluting the signal in primary data. For the small data set, radiomics features outperformed deep features in the classification of benign and malignant tumors. This work highlights the importance of architecture design in transfer learning, and the proposed pipeline is anticipated to provide a reference and inspiration for small data analysis.
PURPOSE: Computed tomography (CT) images can display internal organs of patients and are particularly suitable for preoperative surgical diagnoses. The increasing demands for computer-aided systems in recent years have facilitated the development of many automated algorithms, especially deep convolutional neural networks, to segment organs and tumors or identify diseases from CT images. However, performances of some systems are highly affected by the amount of training data, while the sizes of medical image data sets, especially three-dimensional (3D) data sets, are usually small. This condition limits the application of deep learning. METHODS: In this study, given a practical clinical data set that has 3D CT images of 20 patients with renal carcinoma, we designed a pipeline employing transfer learning to alleviate the detrimental effect of the small sample size. A dual-channel fine segmentation network (FS-Net) was constructed to segment kidney and tumor regions, with 210 publicly available 3D images from a competition employed during the training phase. We also built discriminative classifiers to classify the benign and malignant tumors based on the segmented regions, where both handcrafted and deep features were tested. RESULTS: Our experimental results showed that the Dice values of segmented kidney and tumor regions were 0.9662 and 0.7685, respectively, which were better than those of state-of-the-art methods. The classification model using radiomics features can classify most of the tumors correctly. CONCLUSIONS: The designed FS-Net was demonstrated to be more effective than simply fine-tuning on the practical small size data set given that the model can borrow knowledge from large auxiliary data without diluting the signal in primary data. For the small data set, radiomics features outperformed deep features in the classification of benign and malignant tumors. This work highlights the importance of architecture design in transfer learning, and the proposed pipeline is anticipated to provide a reference and inspiration for small data analysis.
Authors: Thomas De Perrot; Jeremy Hofmeister; Simon Burgermeister; Steve P Martin; Gregoire Feutry; Jacques Klein; Xavier Montet Journal: Eur Radiol Date: 2019-02-12 Impact factor: 5.315
Authors: Bjoern H Menze; B Michael Kelm; Ralf Masuch; Uwe Himmelreich; Peter Bachert; Wolfgang Petrich; Fred A Hamprecht Journal: BMC Bioinformatics Date: 2009-07-10 Impact factor: 3.169