Lesion detection in stroke patients based on intra-modality features on Diffusion MRI

Motivation: Deep learning methods have become increasingly prevalent in the recognition of abnormal brain tissue. However, the reliance on a single modality within datasets often hampers these approaches. Goal(s): We integrate the diffusion sequence derived from a single modality into the model training process. This aims to provide a more comprehensive set of image information, thereby aiding the model in detecting anomalies with greater precision. Approach: We evaluate the performance of the basic U-Net model1 trained on different modalities using a test set for comparison. Results: Experimental findings indicate that dMRI (ADC, TRACE) is more effective in identifying lesions than traditional T1-weighted images. Impact: This study emphasizes the advantages of diffusion sequences (ADC, TRACE) in detecting brain tissue abnormalities while addressing the limitations of single-modal models. It also presents a novel method for developing multi-modal training models from a single modality.