Miao Wu1,2, Xiaoxia Shen3, Can Lai4, Yuqing You4, Zhiyong Zhao1, Dan Wu5. 1. Key Laboratory for Biomedical Engineering of Ministry of Education, College of Biomedical Engineering and Instrumental Science, Zhejiang University, Hang Zhou, China. 2. College of Medical Engineering and Technology, Xinjiang Medical University, Urumqi, China. 3. Department of Neonatal Intensive Care Unit, Children's Hospital, Zhejiang University School of Medicine, Hangzhou, China. 4. Department of Radiology, Children's Hospital, Zhejiang University School of Medicine, Hangzhou, China. 5. Key Laboratory for Biomedical Engineering of Ministry of Education, College of Biomedical Engineering and Instrumental Science, Zhejiang University, Hang Zhou, China. danwu.bme@zju.edu.cn.
Abstract
BACKGROUND: Differentiating acute bilirubin encephalopathy (ABE) from non-ABE in neonates with hyperbilirubinemia (HB) from routine magnetic resonance imaging (MRI) is extremely challenging since both conditions demonstrate similar T1 hyperintensities. To this end, we investigated whether the integration of multimodal MRI from routine clinical scans with deep-learning approaches could improve diagnostic performance. METHODS: A total of 75 neonates with ABE and 75 neonates with HB (non-ABE) were included in the study. Each patient had three types of multimodal images taken, i.e., a T1-weighted image (T1WI), a T2-weighted image (T2WI), and an apparent diffusion coefficient (ADC) map. The three types of MRI contrasts and their combination were fed into two deep convolutional neural networks (CNNs), i.e., ResNet18 and DenseNet201. The performance of CNNs was compared with a traditional statistical method named logistic regression. RESULTS: We demonstrated that diagnostic methods with the multimodal data were better than any of the single-modal data. Both CNN models outperformed the logistic regression method. The best performance was achieved by DenseNet201 with the combination of three modalities of T1WI, T2WI, and ADC, with an accuracy of 0.929 ± 0.042 and an area under the curve (AUC) of 0.991 ± 0.007. CONCLUSIONS: Our study demonstrated that CNN models with multimodal MRI significantly improve the accuracy of diagnosing ABE. IMPACT: We proposed an efficient strategy of detecting ABE in neonates based on multimodal MRI with deep learning, which achieved an accuracy of 0.929 ± 0.042 and an AUC of 0.991 ± 0.007. We demonstrated the advantage of integrating multimodal MRI in detecting ABE in neonates with HB, using deep-learning models. Our strategy of diagnosing ABE using deep-learning techniques with multimodal MRI from routine clinical scans is potentially applicable to clinical practice.
BACKGROUND: Differentiating acute bilirubin encephalopathy (ABE) from non-ABE in neonates with hyperbilirubinemia (HB) from routine magnetic resonance imaging (MRI) is extremely challenging since both conditions demonstrate similar T1 hyperintensities. To this end, we investigated whether the integration of multimodal MRI from routine clinical scans with deep-learning approaches could improve diagnostic performance. METHODS: A total of 75 neonates with ABE and 75 neonates with HB (non-ABE) were included in the study. Each patient had three types of multimodal images taken, i.e., a T1-weighted image (T1WI), a T2-weighted image (T2WI), and an apparent diffusion coefficient (ADC) map. The three types of MRI contrasts and their combination were fed into two deep convolutional neural networks (CNNs), i.e., ResNet18 and DenseNet201. The performance of CNNs was compared with a traditional statistical method named logistic regression. RESULTS: We demonstrated that diagnostic methods with the multimodal data were better than any of the single-modal data. Both CNN models outperformed the logistic regression method. The best performance was achieved by DenseNet201 with the combination of three modalities of T1WI, T2WI, and ADC, with an accuracy of 0.929 ± 0.042 and an area under the curve (AUC) of 0.991 ± 0.007. CONCLUSIONS: Our study demonstrated that CNN models with multimodal MRI significantly improve the accuracy of diagnosing ABE. IMPACT: We proposed an efficient strategy of detecting ABE in neonates based on multimodal MRI with deep learning, which achieved an accuracy of 0.929 ± 0.042 and an AUC of 0.991 ± 0.007. We demonstrated the advantage of integrating multimodal MRI in detecting ABE in neonates with HB, using deep-learning models. Our strategy of diagnosing ABE using deep-learning techniques with multimodal MRI from routine clinical scans is potentially applicable to clinical practice.