An automated hybrid approach via deep learning and radiomics focused on the midbrain and substantia nigra to detect early-stage Parkinson's disease Provisionally Accepted

Hongyi Chen¹ Xueling Liu² Xiao Luo¹ Junyan Fu² Kun Zhou¹ Na Wang² Yuxin Li^2* Daoying Geng^{1, 2*}

• ¹Academy for Engineering and Technology, Fudan University, China
• ²Huashan Hospital, Fudan University, China

Objectives The altered neuromelanin in substantia nigra pars compacta (SNpc) is a valuable biomarker in the detection of early-stage Parkinson's disease (EPD). Diagnosis via visual inspection or single radiomics based method is challenging. Thus, we proposed a novel hybrid model that integrates radiomics and deep learning methodologies to automatically detect EPD based on neuromelanin-sensitive MRI, namely short-echo-time Magnitude(setMag) reconstructed from quantitative susceptibility mapping (QSM).
Methods In our study, we collected QSM images including 73 EPD patients and 65 healthy controls, which were stratified into training-validation and independent test sets with an 8:2 ratio. Twenty-four participants from another center were included as the external validation set. Our framework began with the detection of the brainstem utilizing YOLO-v5. Subsequently, a modified LeNet was applied to obtain deep learning features. Meanwhile, 1781 radiomics features were extracted, and 10 features were retained after filtering. Finally, the classified models based on radiomics features, deep learning features, and the hybrid of both were established through machine learning algorithms, respectively. The performance was mainly evaluated using accuracy, net reclassification improvement (NRI), and integrated discrimination improvement (IDI). The saliency map was used to visualize the model.
Results The hybrid feature-based support vector machine (SVM) model showed the best performance, achieving ACC of 96.3% and 95.8% in the independent test set and external validation set, respectively. The model established by hybrid features outperformed the one radiomics feature-based (NRI: 0.245, IDI: 0.112). Furthermore, the saliency map showed that the bilateral "swallow tail" sign region was significant for classification.
Conclusion The integration of deep learning and radiomic features presents a potent strategy for the computer-aided diagnosis of EPD. This study not only validates the accuracy of our proposed model but also underscores its interpretability, evidenced by differential significance across various anatomical sites.