Contact: Yunan Wu
To implement a deep learning algorithm to automatically identify a patient as having an emergent finding (at least one of 21 different conditions) on a head CT study in order to create an efficient workflow in the reading room.
Materials and Methods
A total of 1405 CT scans were acquired from 997 subjects. Scan-based, whole brain labels were determined from the radiology reports and used as the gold standard to generate the patient’s condition; emergent or non-emergent. Five-fold cross-validation was used, including 80% as the training dataset and 20% as the validation dataset. For each scan, three window levels were applied to enhance the display of the brain, soft tissues and stroke using the pair of window and center of [80,40], [400,80] and [30,30] respectively. The concatenated inputs were normalized into the range of 0 to 1 and rescaled to the same volume size (48, 256, 256). A sequence to sequence model was designed to make the classification and generate slice-level attention weights. Each CT slice was embedded into a shared 2D convolutional neural network (CNN) to extract feature vectors, which were passed into an attention layer to get a weight for each slice. Then, weighted average feature vectors were computed for each CT scan and fed into two fully connected layers, followed by the sigmoid activation function to predict the patient’s status. A gradient back-propagation-based visualization method (Grad-CAM) was applied to the volume to map the location of the emergent findings.
The sequence CNN model was able to identify the presence of any of the 21 conditions in the emergent category averaged across all five-folds with an overall accuracy of 0.848 ± 0.003, sensitivity of 0.862 ± 0.003, and ROC-AUC score of 0.872 ± 0.002. The attention weights for each scan successfully identified the key axial slices and the Grad-CAM mappings from these slices demonstrated the discriminative locations used by the model.
Our current sequence CNN model can successfully detect head CT emergencies and achieve comparable performance with current literature with only the volume-level labels. The visual interpretation of the model provided an important way to alert radiologists of unexpected brain emergencies. In the future, we expect to scale up our experiment with the addition of 33,000 Head CT Scans and their corresponding radiology reports.
Teams: Yunan Wu, Amit Sanjay Adate, Shamal Shashi Lalvani, Michael Iorga, Todd B Parrish, Aggelos K. Katsaggelos, Virginia Hill
Slice-based Attention Weights on a Scan & Slice Visualization Mappings