A Novel Approach for Aphasia Evaluation based on ROI-based Features from Structural Magnetic Resonance Image

• Published in: IEEE journal of biomedical and health informatics 2024-11, Vol.PP, p.1-13
• Main Authors: Dan, Ying, Cai, Aiqun, Ma, Jiaxin, Zhong, Yuming, Mahmoud, Seedahmed S., Fang, Qiang
• Format: Article
• Language: English
• Subjects: Accuracy; Aphasia; Computer-aided Diagnosis; Costs; Data models; Feature extraction; Lesions; Machine Learning; Magnetic resonance imaging; Medical diagnostic imaging; Post-stroke aphasia; Predictive models; Radiomics
• ISSN: 2168-2194; 2168-2208; 2168-2208
• DOI: 10.1109/JBHI.2024.3492072

Aphasia, affecting one-third of stroke survivors, impairs language comprehension and speech production, leading to challenges in daily interactions, social isolation, and economic losses. Assessing aphasia is crucial for effective rehabilitation and recovery in patients. However, the conventional behavioral-based evaluation, reliant on speech pathologists, is susceptible to individual variability, resulting in high labor costs, time-consuming processes, and low robustness. To address these limitations, this study introduces a novel evaluation method based on medical image processing and artificial intelligence. Magnetic resonance imaging (MRI) provides exceptional spatial resolution while mitigating the impact of individual variability. The image processing techniques were employed to extract pathological features, specifically region-of-interest (ROI)-based features. Subsequently, the evaluation models were trained using ROI-based features which initially identify the occurrence of aphasia and then categorize the type of aphasia, aiding clinicians in tailoring treatment to various therapeutic approaches and intensities. The evaluation models also predict the severity and generate scores for four types of language function: spontaneous speech, auditory comprehension, naming, and repetition. Both aphasia occurrence detection and aphasia type classification attain impressive accuracy rates of 100.00 \pm 0.00% and 85.00 \pm 13.23%, respectively. The severity prediction yields the lowest root mean square error (RMSE) of 17.03 \pm 2.75, while the assessment of four language functions achieves the best RMSE of 1.27 \pm 0.82. Utilising the advantages of a medical imaging-based automation approach, the proposed aphasia evaluation method provides a comprehensive procedure and generates rather accurate results. Hence it could assist the aphasia rehabilitation and substantially reduce clinicians' workload.