Abstracts

RETURN TO ABSTRACT LISTING


E1965. Utilization of a Novel Deep Learning Model for Diagnosing Intracranial Aneurysms Using MRA Source and MIP Data
Authors
  1. Richard Lee; Santa Clara Valley Medical Center
  2. Pradnya Patel; Santa Clara University
  3. Mahesh Patel; Santa Clara Valley Medical Center
  4. Yuling Yan; Santa Clara University
  5. Young Kang; Santa Clara Valley Medical Center
Objective:
Evaluation of a novel deep learning algorithm for detection of intracranial aneurysms using data input from MRA source images and MRA MIPs.

Materials and Methods:
Contiguous MRA images reported by radiologists in a community hospital setting to contain aneurysms were obtained from between January 2017 through December 2019. The images were divided into two data sets: training data set and test data set. A novel deep learning model using a convolutional neural network (CNN) was constructed, comprising of 11 layers: 3 layers of convolution, 3 max pool layers, 3 fully connected layers, a softmax layer, and an output. The CNN model was trained and validated on the CBIS-DDSM (Curated Breast Imaging Subset of the Digital Database for Screening Mammography) dataset. The CBIS-DDSM is an open source dataset comprising of 2620 scanned mammography films, containing 10239 images from 1566 participants. which achieved an accuracy of 93.4%. Transfer learning was then used to retrain the model further for detecting aneurysms with our MRA dataset, which contains 29024 images for normal MRA and 25245 images for MRA with aneurysms collected from 200 patients in a community setting, out of which 100 are healthy and 100 have aneurysm.

Results:
The training data set, which provided training and validation data, included 140 examinations (70 with aneurysms and 70 without). Test data included 60 (30 with aneurysms and 30 without). Model achieved an accuracy of 76.04%, sensitivity of 74.89%, and specificity of 78.52%.

Conclusion:
By establishing usefulness of this novel deep learning model in incorporating data from MRA source and MIP datasets, we can use it to increase accuracy and efficiency of future diagnosis of intracranial aneurysms in the community setting. Early and accurate detection of unruptured aneurysms may significantly improve clinical outcomes. This novel deep learning model, with incorporation of both MRA source and MRA MIP data, can use fewer training data sets. Our model resulted in sensitivity and specificity less than previously-reported studies although may still aid in detecting aneurysms in the community clinical setting.