At present, medical image fusion depends on the source images for generating results because the dataset lacks ground truth data. This constraint can result in less effective fusion outcomes, potentially causing valuable information generated during the training process to be overlooked or discarded. Additionally, numerous current fusion techniques that utilize convolutional neural networks rely on local receptive fields, which results in an inability to fully capture global information from the domain. To overcome these challenges, we introduce a medical image fusion framework called IUFusion, which relies on information units. Firstly, we design an information unit to obtain the weights of the superior results produced during training, which indicate the degree of feature retention of the intermediate results. These weights are then introduced into our loss function, allowing the intermediate results to assist the source image in supervising the training process. Furthermore, we design a multi-channel attention information unit that incorporates the attention mechanism into the channel and utilizes it to synthesize global contextual information about multimodal features. Our method surpasses the most advanced fusion methods, as evidenced by both qualitative and quantitative experiments.