The paper lacks a thorough discussion of the limitations of the proposed method. While it mentions some limitations in the ablation study, it does not delve into the potential drawbacks of using transformers for medical image segmentation. A more detailed discussion of the limitations, such as computational cost, data requirements, and potential biases, would strengthen the paper.

The paper's evaluation is primarily focused on two datasets, BTCV and MSD. Evaluating the method on a wider range of datasets, particularly those with different imaging modalities and anatomical structures, would provide a more comprehensive assessment of its generalizability.

The paper compares UNETR against several existing methods, but the comparison is not always fair. For example, in the BTCV benchmark, the comparison is against methods that use different training data or different evaluation protocols. A more rigorous comparison, using the same data and evaluation metrics for all methods, would strengthen the results.