sequence, and they had to create a single model capable of handling multiple input combinations and output scenarios. They also needed to quantify the contribution of each input sequence. “To solve this, we provide a taskspecific weighted average,” Luyi explains. “When we want to quantify the contribution, we think about the weighted average because if they contribute more, they have higher weight. We input the module with a combination of the models. For example, if we have T1, T2, and Flair and want to output T1Gd, we make a zero-one code to present this conditional input. For example, we can build it as 1011 and 0100 to present this input and output condition. With this condition, we can use a fully connected layer to predict a weight. This weight can be used as a weighted average for the sequence fusion. We can train the model and learn the weighted average based on the fully connected layer.” Also, Luyi and Tianyu designed a task-specific attention module. This module employs conditional channel and spatial attention to generate residual attentional fused features to refine the synthesis performance. The fusion feature is divided into a directly weighted average from the input sequence features and an attention model to generate residual fused features. “We use the weighted average feature as a basic feature,” Luyi continues. “We can directly reconstruct the image from this feature. Also, we build the attention 13 DAILY MICCAI Wednesday Luyi Han
RkJQdWJsaXNoZXIy NTc3NzU=