In this paper, Luyi and joint first author Tianyu Zhang contend with two fundamental questions in multi-sequence MRI synthesis: how to quantify the contribution of different input sequences in synthesizing a missing sequence and how to estimate the quality of the generated image at the pixel level. Medical professionals rely on multiple MRI sequences for diagnosis and prognosis in many clinical scenarios. However, certain sequences are sometimes missing due to various factors and need to be synthesized. Recent works have primarily overlooked a crucial aspect: understanding which input sequences contribute more significantly to the synthesis process. Also, many works provide uncertainty maps presenting the uncertainty of output results and attention-based techniques that output attention maps to analyze regions of uncertainty in images. However, uncertainty-based methods require multi-time outputs to calculate standard deviation while attentional-based methods have limitations in terms of the low quality of the visualizations. “These two questions concern the synthesis model’s explainability and reliability,” Luyi tells us. “Unlike natural image analysis, our medical image analysis model must be more reliable and explainable because we need to use it in a clinical environment.” The researchers faced several challenges during their work. They used the BraTS2021 dataset, which includes four sequences: T1, T1Gd, T2, and Flair. There are many input combinations to synthesize a missing 12 DAILY MICCAI Wednesday Poster Presentation An Explainable Deep Framework: Towards Task-Specific Fusion for Multi-to-One MRI Synthesis Luyi Han is a PhD student at Radboud University Medical Centre and the Netherlands Cancer Institute. He speaks to us about his work on multi-sequence MRI synthesis ahead of his poster this afternoon.
RkJQdWJsaXNoZXIy NTc3NzU=