Performance not comparable to the co-responding research paper

[EthicalAccount]

I was not able to reproduce the results that were provided in the research paper.

Using Mamba pretrained model, the results were not satisfactory and were significantly lower as compared to the proposed values.

These results were of Mamba using pre-trained checkpoint weight.

I also tried using simple UNet architecture and it too had performance degradation, while also simultaneously performing better than Mamba-UNet

The dataset used was ACDC and segmentation task was performed upon it.

It also took significantly longer for mamba tiny to train (around 20 hours)
Specifications
CPU: Intel i9-11900K
RAM: 64 GB DDR4
GPU: RTX 3070 8GB

Everything ran on default settings with the parameters cross-validated as provided in the research paper.