Data Modeling
Mathematical and Computational Data Modeling
0
Citations
4.5k
Views
29
Articles
Volume 1 Number 1 2020
RESEARCH ARTICLE (Open Access)
Automated Liver Segmentation from T1-Weighted Abdominal MRI Using UNet++ with DenseNet Backbones: A Comparative Enhancement-Driven Approach
Kamruzzaman Mithu 1*
Data Modeling 1 (1) 1-11 https://doi.org/10.25163/data.1110843
Submitted: 17 January 2020 Revised: 04 March 2020 Accepted: 13 March 2020 Published: 15 March 2020
Abstract
Liver segmentation is, in many ways, the quiet first step that everything else in AI-assisted hepatic diagnosis depends on — yet doing it reliably from magnetic resonance imaging has proven harder than doing it from CT, largely because MRI lacks a standardized intensity scale comparable to Hounsfield units. This study set out to examine whether that gap could be narrowed, at least for T1-weighted abdominal scans, through a combination of architectural choice and image enhancement. Working with 647 T1-weighted slices drawn from 20 healthy participants in the CHAOS challenge dataset, we evaluated the UNet++ segmentation architecture paired with DenseNet encoder-decoder backbones of varying depth, trained on three distinct input conditions: unmodified slices, gamma-corrected slices, and three-channel RGB composites built from contrast-enhancement transforms. A five-fold cross-validation protocol, with consistent training hyperparameters across conditions, was used to keep the comparison fair. The gamma-corrected condition, paired with a DenseNet-201 backbone, came out ahead of the alternatives, reaching a Dice Similarity Coefficient of 91.91% and an Intersection over Union of 89.27% — figures that place this configuration favorably against prior MRI-based segmentation work on the same dataset, while still falling short of the near-ceiling performance CT-based pipelines typically achieve. Predicted masks were also visually smoother and more anatomically consistent under this configuration than under the alternative enhancement strategies. Taken together, these findings suggest that targeted, relatively simple intensity enhancement — rather than more elaborate multi-channel composites — may be a practical, underused lever for closing the MRI-CT performance gap in automated liver segmentation.
Keywords: Liver segmentation; UNet++; DenseNet; T1-weighted MRI; image enhancement
References
Alves, F. C., Brito, J., et al. (2007). Liver haemangioma: Common and uncommon findings and how to improve the differential diagnosis. European Radiology, 17.
Hu, P., Wu, F., Peng, J., Liang, P., & Kong, D. (2016). Automatic 3D liver segmentation based on deep learning and globally optimized surface evolution. Physics in Medicine & Biology, 61, 8676–8698.
Kavur, A. E., Gezer, N. S., Baris, M., Aslan, S., Conze, P. H., Groza, V., Pham, D. D., Chatterjee, S., Ernst, P., Özkan, S., & Baydar, B. (2020). CHAOS challenge – Combined (CT-MR) healthy abdominal organ segmentation. Medical Image Analysis, 69, 101950. https://doi.org/10.1016/j.media.2020.101950
Kim, K., & Chun, J. (2020). A new hyperparameter of Hounsfield unit range in liver segmentation. Journal of Internet Computing and Services, 21(3).
Liu, Z., Song, Y., et al. (2019). Liver CT sequence segmentation based on improved U-Net and graph cut. Expert Systems with Applications, 126.
Mostafa, A., Hassanien, A. E., Houseni, M., & Hefny, H. (2017). Liver segmentation in MRI images based on whale optimization algorithm. Multimedia Tools and Applications, 76, 24931–24954.
Mulay, S., Deepika, G., Jeevakala, S., Ram, K., & Sivaprakasam, M. (2019). Liver segmentation from multimodal images using HED-Mask R-CNN. In Multiscale Multimodal Medical Imaging: First International Workshop, MMMI 2019, held in conjunction with MICCAI 2019 (Lecture Notes in Computer Science, Vol. 11977, pp. 68–75). Springer.
Peng, J., Wang, Y., et al. (2014). Liver segmentation with constrained convex variational model. Pattern Recognition Letters, 43.
Tang, X., Jafargholi Rangraz, E., Coudyzer, W., Bertels, J., Robben, D., Schramm, G., Deckers, W., Maleux, G., Baete, K., Verslype, C., et al. (2020). Whole liver segmentation based on deep learning and manual adjustment for clinical use in SIRT. European Journal of Nuclear Medicine and Molecular Imaging, 47, 2742–2752.
1
Save
Save
0
Citation
Citation
7
View
View
0
Share
Share