An efficient pyramid scene parsing network with multi-scale feature fusion for liver segmentation in magnetic resonance imaging.

Background: Recent lifestyle changes have led to an increase in the number of liver disease cases, making accurate liver segmentation increasingly important for clinical applications. However, manual segmentation is time-consuming, operator-dependent, and can vary between experts, creating a need for reliable automated approaches. Methods: In this study, a PSP-EffB0-MSFF model is proposed for 2D liver segmentation from abdominal MRI images. EfficientNetB0 replaces the ResNet50 backbone in PSPNet to reduce computational cost, and a multi-scale feature fusion module is incorporated into the skip connections to improve feature representation. The model is evaluated using two datasets: T1-weighted images from DLDS and T2-weighted images from CirrMRI600+. Results: On the DLDS dataset, the model achieves an intersection over union of 0.905 ± 0.038 and a Dice score of 0.913 ± 0.09, along with a Hausdorff distance of 7.31 ± 3.91 and an average sym metric surface distance of 2.66 ± 3.06. On the CirrMRI600+ dataset, it achieves an intersection over union of 0.86 ± 0.01 and a Dice score of 0.90 ± 0.02, with a Hausdorff distance of 6.20 ± 0.60 and an average symmetric surface distance of 9.80 1 1.50 at the patient level. The model requires 14.91 GFLOPs. Conclusion: Overall, the proposed PSP-EffB0-MSFF model provides reliable segmentation results on CirrMRI600+ and shows consistent performance on DLDS under the current experimental setup.