Embedded Files
Whole slide imaging (WSI) is an emerging imaging technology that has the potential to revolutionize the field of digital pathology, but its acceptance in computer-aided diagnostics is hindered by a relative scarcity of labeled data sets compared to those obtained using traditional imaging technologies. Exhaustive manual labeling and annotation are significantly time-consuming due to the sheer size of the images, and require technical expertise not readily available among the general population. The use of artificial intelligence for such tasks, on the other hand, suffers from low confidence in results and a lack of transparency as perceived by the medical community.
WSIs capture incredible detail, but their size and complexity
make them hard to analyze algorithmically.
This work presents an unsupervised segmentation algorithm which draws on sparse representation theory and
cluster analysis to provide annotations with greater algorithmic transparency than traditional neural networks.
This work presents an unsupervised segmentation algorithm which draws on sparse representation theory and
cluster analysis to provide annotations with greater algorithmic transparency than traditional neural networks.
The algorithm learns a sparse dictionary representation for all patches in an image...
... and partitions the dictionary into multiple sub-dictionaries based on atom utilization.
The extracted image patches are then reconstructed using each sub-dictionary individually and classified per the
best-performing dictionary, with performance gauged by a weighted sparsity and reconstruction error metric.
The extracted image patches are then reconstructed using each sub-dictionary individually and classified per the
best-performing dictionary, with performance gauged by a weighted sparsity and reconstruction error metric.
Relevant Works
Relevant Works
[T01] Hierarchical Semantic Segmentation of Histopathological Whole Slide Images by Means of Sparse Dictionary Learning
[T01] Hierarchical Semantic Segmentation of Histopathological Whole Slide Images by Means of Sparse Dictionary Learning
A. S. Levin
M.Eng. Thesis, The Cooper Union, 2021
🔗 [Publication] · [PDF]
[S04] Defense Hierarchical Semantic Segmentation of Histopathological Whole Slide Images by Means of Sparse Dictionary Learning
[S04] Defense Hierarchical Semantic Segmentation of Histopathological Whole Slide Images by Means of Sparse Dictionary Learning
A. S. Levin
M.Eng. Thesis Defense, The Cooper Union, 2021
🔗 [Slides]
Code Repository
Code Repository
The full implementation is available on GitHub:
GitHub – WSI Sparse Segmentation
Dataset
Dataset
The dataset of annotated glomerulus regions from mouse kidney WSIs is publicly available at:
Good Glomeruli Dataset
Acknowledgements
Acknowledgements
This work was conducted as part of my Master’s of Engineering research at The Cooper Union. I am grateful to Prof. Fred Fontaine, who has been my advisor and mentor not only on this thesis, but throughout all my years at Cooper. I would also like to thank Dr. Avi Rosenberg of The Johns Hopkins University School of Medicine, Dr. Pinaki Sarder of the University at Buffalo, and their respective teams for their guidance in this project.
Page updated
Google Sites
Report abuse