E1761. Segmentation of Intracerebral Hemorrhage Using a U-Net Architecture With 2-Channel Input
  1. David Wang; University of Cincinnati
  2. Lily Wang; University of Cincinnati
  3. Brady Williamson; University of Cincinnati
  4. Vivek Khandwala; University of Cincinnati
  5. Tyler Behymer; University of Cincinnati
  6. Daniel Woo; University of Cincinnati
  7. Achala Vagal; University of Cincinnati
Prediction of outcomes of intracerebral hemorrhage (ICH) is critical and volumetric measurement of the ICH is often the first step. Applying traditional machine learning methods or incorporating sequential or parallel deep learning models has greater predictive value in accurate segmentation of the ICH. This project aims to segment ICH in head CT in patients with thalamocapsular hemorrhage, with or without intraventricular hemorrhage (IVH).

Materials and Methods:
We utilized the National Institutes of Health funded, multicenter Ethnic/Racial Variations of Intracerebral Hemorrhage (ERICH) study database and included a subset of patients with right thalamocapsular hemorrhage (1). We selected the initial non-contrast computed tomography (CT) brain of each patient. The training set contained 273 whole brain CT studies, the validation and test sets contained 20 CT studies each, with the sets being mutually exclusive. Images were stratified by center prior to randomization into validation and test sets. The images were deidentified manually and reshaped to 512 x 512 x 30 with linear interpolation. Window values were constrained to 0-200 Hounsfield units (HU) and 30-60HU, remapped to a range of [0, 1] and added to the channel layer of the input (shape 512 x 512 x 30 x 2, channels last). Segmentation masks were manually drawn by the first author. Output for the model had shape 512 x 512 x 30 x 1, as binary segmentation masks. The model was written in Tensorflow 2.0 utilizing a U-net architecture, which was loosely based on the idea of a hybrid 2/3D model (2). The first 3D convolution in the encoding arm used a kernel size of 3 x 3 x 3 and stride of 1. Subsequent convolutions were all 2D. Each convolution block in the model followed the format: conv3d > batch normalization > drop out > leaky ReLU. Drop out was set at 0.05 and convolutions used He normal initialization. The final layer used a sigmoid activation. Data was augmented with random left-right flips during training on a Nvidia RTX 2080 Ti. Learning rate used was 1e-4, with an Adam optimizer. Batch size of 1 was used due to hardware limitations. Loss was defined using a modified soft dice loss (3) with smoothing of 1. The learning rate was decreased by a factor of 0.5 on validation loss plateau, with a patience of 5 epochs. Training was terminated on plateau of validation loss, with a patience of 50 epochs. The model with the lowest validation loss was saved.

The training, validation and test sets contain 66%, 55% and 85% IVH, respectively. On the validation set, the model had a best loss of 0.154 and a dice score of 0.821. On the test set, the model achieved a loss of 0.149 and a dice score of 0.824.

This study demonstrates the feasibility of segmenting ICH from IVH in CTs from multiple centers, using a 2-channel input. Further training data is needed to encapsulate different hemorrhage patterns. Future directions include the use of other model architectures, such as fully 3D convolutions, and validating our method in different datasets.