Search

Medical image-based artificial intelligence (AI) systems have shown great potential in assisting disease diagnosis and treatment. However, the challenges, such as data silos, privacy security and standardization, seriously impedes the application of AI in disease diagnosis and treatment. By integrating federated learning technology and FAIR data principle, it is possible to resolve the aforementioned obstacles. Then, it is able to maximize the value of multicenter data to develop a more efficient and accurate disease diagnosis and treatment AI systems, and promote the clinical application of medical image-based AI systems in the field of disease diagnosis and treatment.

Objective: To evaluate the feasibility, robustness and reproducibility of radiomics features derived from lung diffusion-weighted imaging (DWI). Methods: Thirty patients with pulmonary nodules/masses who underwent magnetic resonance imaging examination in the Department of Radiology, the First Affiliated Hospital of Guangzhou Medical University, from January 4 2019 to May 5 2019, including 16 males and 14 females, aged from 27 to 69 (57±11) years, were prospectively collected. Planar echo imaging (EPI) -DWI and fast spin-echo (TSE) -DWI scans were performed under free-breathing conditions. Each scan was repeated at an interval of 5 minutes, and the corresponding apparent diffusion coefficient (ADC) maps were reconstructed. Each DWI and ADC sequence (a total of eight groups of images) were manually segmented by two radiologists, and a total of 396 radiomics features in 6 categories were extracted from each group of images. Consistency correlation coefficient (CCC) and dynamic range (DR) were used to evaluate the robustness of features between two scans, and stable features were defined as both CCC values and DR values ≥0.85. Intra-observer and interobserver reproducibility were evaluated by intra-group correlation coefficient (ICC), and ICC values≥0.75 was considered to be good reproducibility. Results: Regardless of EPI or TSE technique, the number of robust features extracted fromDWI (TSE: n =197, EPI: n =169) were higher than that of the corresponding ADC (TSE: n =126, EPI: n =148). The proportion of robust features of TSE-DWI、EPI-DWI、TSE-ADC、EPI-ADC was 49.7% (197/396), 42.7% (169/396), 31.8% (126/396) and 37.4% (148/396), respectively. Of the 396 features, 54 (13.6%) of them demonstrated great robustness (CCC and DR≥0.85) and interobserver and interobserver reproducibility (ICC≥0.75) across all sequences. Conclusions: Radiomics features derived from lung DWI showed robustness and reproducibility. Different sequences and different feature clusters have different proportions of stable features, and some features have good robustness and reproducibility between different scans, different observers, and even different sequences.

Objective: To explore the value of radiomics model based on dynamic contrast-enhanced magnetic resonance imaging (MRI) in differentiation fat-poor angiomyolipoma (fp-AML) from alpha-fetoprotein-negative hepatocellular carcinoma (n-HCC) in the background of non-cirrhotic liver. Methods: The complete data of 121 patients from Guangdong Provincial People's Hospital, Zhongshan Hospital Affiliated to Fudan University and Sun Yat-sen University Cancer Center with hepatic fp-AML and n-HCC confirmed by pathology from October 2010 to July 2020 were retrospectively analyzed. Among them, 75 were males and 46 were females, aged from 23 to 80 (55±12) years. A total of 93 patients from Zhongshan Hospital Affiliated to Fudan University were divided into the training cohort ( n =75) and internal test cohort ( n =18) according to entry time, and the patients of other 2 hospitals were divided into external test cohort ( n =28). The radiomics features were extracted from the preoperative triple-phase contrast-enhanced images. The feature selection algorithm based on Joint Mutual Information Maximisation (JMIM) was used to extract the optimal feature subset, and support vector machine (SVM) was used to build the radiomics model. The diagnostic performance of radiomics model was evaluated using the receiver operating characteristic (ROC) curve, and was compared with that of two radiologists. Results: In the internal cohort, the area under the curve (AUC) for the differential diagnosis between fp-AML and n-HCC of the radiomics model was 0.819 (with an accuracy of 72.2%), outperforming than radiologist 1 with 10 years of diagnostic experience (AUC=0.542, P =0.029) and radiologist 2 with 2 years of diagnostic experience (AUC=0.375, P =0.004). In the external cohort, the AUC of the radiomics model was 0.772 (with and accuracy of 71.4%), which was comparable to that of radiologist 1 (AUC=0.661, P =0.442) and better than that of radiologist 2 (AUC=0.400, P =0.008). Conclusion: The radiomics model based on dynamic contrast-enhanced MRI is of high accuracy for preoperatively differentiating hepatic fp-AML from n-HCC in the noncirrhotic liver.