An Omicsmap-Driven Deep Learning Model for Transcranial Sonography–Based Parkinson's Disease Assessment
Abstract
Purpose
The global prevalence of Parkinson’s disease (PD) is rising in tandem with a rapidly aging population. Transcranial sonography (TCS) has emerged as a promising tool for early diagnosis, as it is radiation-free, highly accessible, and cost-effective. Specifically, TCS enables the evaluation of substantia nigra (SN) echogenicity in the midbrain—a well-established and reliable biomarker for PD. However, current clinical reliance on manual evaluation is limited by operator subjectivity, highlighting an urgent need for AI-driven TCS interpretation systems. This study aims to advance the field by leveraging image-based representations of TCS radiomics data integrated with clinical features.
Methods
TCS images and clinical data were retrospectively collected from 324 subjects (PD:103, HC: 221) who were either suspected or diagnosed with PD. Clinical data include patient demographics, cognitive function, and symptom severity. Experienced sonographers delineated SN on TCS images. All pre-processed TCS images underwent radiomics feature extraction. Each subject’s combined clinical and radiomic feature vector (with and without feature selection) is transformed into a 2D topographic map, called OmicsMap. CNNs includes ConvNext, EfficientNet and ResNet were then applied separately by OmicsMap construction. Model performance was evaluated on the independent testing cohort using AUC, Accuracy, Balanced Accuracy and F1-score.
Results
ConvNeXt achieved the best performance (AUC = 0.813, ACC= 0.744, BACC= 0.743 and F1=0.649), outperforming ResNet (AUC = 0.715, ACC= 0.700, BACC= 0.664 and F1=0.661) and EfficientNet (AUC = 0.673, ACC= 0.620, BACC= 0.651 and F1=0.592) in distinguishing between PD and HC, suggesting its superiority in PD assessment by leveraging OmicsMap-driven ConvNeXt based on feature selection. Additionally, OmicsMap-based models outperformed those without OmicsMap in PD assessment.
Conclusion
The encouraging findings of this study demonstrated the promising potential of integrating ultrasomics features, clinical predictors using OmicsMap-driven model for PD assessment.