Please enable JavaScript.
Coggle requires JavaScript to display documents.
Hemorrhage Detection, Grewal, M., Srivastava, M. M., Kumar, P., and …
Hemorrhage Detection
Traumatic Brain Injury
One cause of hemorrhage
Fast detection is required
Delays in detection could cause sever complications
Requires Computed Tomography scans
Requires evaluation of medical experts
Time consuming to manually analyze
Automated Detection
Faster inference
Potential to help more people
Use deep learning methods
Capable to achieve radiologist level accuracy
Modelling 3D CT scans
Classification
Sequence Labeling
Benchmark to human experts
Results and Discussion
RadNet outperformed other CNNS
Compared to baseline DenseNet
DenseNet with attention maps yielded better recall scores
Test set compared with results of expert radiologists
Achieved high scores
Highest on Recall scores
Radiologists still proved to be better
Test set annotations from independent results from other experts
Sensitivity metric is important
Network Architecture
RadNet
Baseline CNN
Used for slice level features
LSTM
Bidirectional learning
Model 3D images as sequences
Use of attention mechanisms
Hard Attention
Focusing on certain parts of an image
Soft Attention
Focus on certain features
Materials and Methods
Dataset of brain CT scans
Annotations based on slices
Contour segmentation
Architecture
40 layer DenseNet
Adds auxiliary functions
Outputs segmented regions of an image
Feeds output of CNN to LSTm
Grewal, M., Srivastava, M. M., Kumar, P., and
Varadarajan, S. Radnet: Radiologist level accuracy
using deep learning for hemorrhage detection in ct
scans, 2018.