Deep Learning for Brain Tumor Segmentation
Author Information
Author(s): Yoo Jay J., Namdar Khashayar, Khalvati Farzad
Primary Institution: University of Toronto
Hypothesis
Superpixels can be leveraged to acquire additional contextual information, thereby improving weakly supervised segmentation performance.
Conclusion
The proposed combination of deep superpixel generation, deep superpixel clustering, and the incorporation of undersegmented seeds as a loss term improves weakly supervised segmentation.
Supporting Evidence
- The proposed method achieved a mean Dice coefficient of 0.745 on the BraTS 2023 dataset.
- The method outperformed state-of-the-art baselines including CAM-S and SAM.
- The study utilized 2D slices of MR images for training and evaluation.
Takeaway
This study created a smart way to find brain tumors in pictures by using computer programs that learn from examples, even when there aren't many labeled pictures available.
Methodology
The study used a deep superpixel generation model and a clustering model trained simultaneously to segment brain tumors from MR images using weakly supervised learning.
Potential Biases
The reliance on superpixels based on pixel intensity may limit the method's applicability to non-focal regions of interest.
Limitations
The method relies on the accuracy of localization seeds, and poor seeds can lead to poor segmentations; it also cannot be trained end-to-end.
Participant Demographics
The dataset included 369 patients with brain tumors from the BraTS 2020 dataset.
Digital Object Identifier (DOI)
Want to read the original?
Access the complete publication on the publisher's website