Keywords:
Artificial Intelligence, CNS, CT, Computer Applications-Detection, diagnosis, Computer Applications-General, Trauma, Aneurysms
Authors:
W. Ding, J. W. Luo, J. J. R. Chong; Montreal, QC/CA
DOI:
10.26044/ecr2019/C-3184
Methods and materials
Study Population:
A retrospective case-control review was performed of all adult CT head examinations on PACS from January 2006 to December 2017 at two academic tertiary care hospitals.
In order for a case to be included,
single-slice images had to contain a single focus of SAH,
to aid in focusing training on a single finding,
given the weak localization labelling provided.
Images containing multi-compartmental bleeds were excluded as these were hypothesized to possibly compromise training of the network.
Control images were derived from independent CT head examinations without intracranial hemorrhage,
specified via a normal impression or conclusion from final clinical radiology reports.
Image Pre-Processing & Training Labels:
The standard diagnostic complete DICOM studies are exported.
The primary standard window axial series was isolated and converted to a downsampled 256x256px input image.
During this conversion,
standard filter manufacturer window width/window level settings are maintained.
Individual slice images were anonymized as per standard protocols.
Individual slice images were then annotated and reviewed by two consultant radiologists for the presence or absence of intracranial bleed as well as subarachnoid bleed versus multi-compartmental bleed specifically.
Neural Network Configuration:
Utilizing a ImageNet-based transfer learning methodology with image augmentation,
a ResNet-50 convolutional neural network was fine-tuned to image classify between the two desired output classes,
utilizing a 75:25 training/validation split (200 Epochs; ADAM optimizer,
LR=0.001).
Given the limited minority class images available for training,
splits were performed with per-study (i.e.
CT scan) isolation to ensure isolation of patient studies between the training and validation groups.
Training and validation runs were repeated using a 5-cross fold validation.
The neural network implementation used a Python/Tensorflow+Keras framework for neural network implementation.
Given the nature of the desired findings,
classification of intracranial bleeds,
standard training image generator augmentation policies were used of standard affine transformations (e.g.
zoom,
horizontal flip,
shear,
rotation).
Training was performed on a deep learning workstation with a single Nvidia Titan X Pascal GPU (Santa Clara,
CA).
Neural Network Objective and Subjective Validation:
Objective validation of the trained CNN was performed using ROC curve analysis of the inferred predictions on the withheld validation group using Area Under the Curve (AUC) measures [8].
Given the limited dataset,
further subjective evaluation of the validation set predictions was performed with localization heat-map analysis,
using a combination of both Salience Maps and Gradient-weighted Class Activation Mapping (Grad-CAM) [9].