Federated learning (FL) is increasingly emerging as a collaborative learning paradigm, particularly for multiple medical units. By enabling model training across distributed datasets without sharing raw data, FL addresses critical privacy concerns across different medical units. Most FL methods only considered intramodal heterogeneity within one client, ignoring the challenge of data heterogeneity in the presence of missing modalities between multiple clients. In real-world scenarios, missing modalities, where some centers lack certain data types, and dominant modalities, where specific data types disproportionately influence training, can result in the emergence of dominant clients. These dominant clients can easily lead to a suboptimal final model without the dynamic aggregation weights adjustment. To address this challenge, we propose a novel Federated Discrepancy Correction Reweight Aggregation (FedDCRA), designed to mitigate the disproportionate influence of dominant clients. To measure the aggregation weight, we calculate the discrepancy between each local model and the global model. Then we utilize the singular value decomposition and weighted re-aggregation to obtain the corrected discrepancies with noise removed. Finally, we use the discrepancy before and after correction to generate aggregation weights. To accurately calibrate the weights to the dominant client, we use the ratio of each client’s current loss to its historical loss, which reflects the client’s convergence rate, to adjust the aggregation weights and adaptively tune the learning rate of each client for subsequent rounds. Extensive experimental results on BRATS2020 and BRATS2018 demonstrate the superiority of our method.