Convolutional neural networks to identify malformations of cortical development: A feasibility study
Iván Sánchez Fernández1, Edward Yang2, Marta Amengual-Gual3
1Boston Medical Center, Boston University School of Medicine, Boston, MA USA.
Seizure
|June 15, 2021
Summary
Deep learning models accurately detect malformations of cortical development (MCD) using MRI scans. Convolutional neural networks (CNNs) offer a fully automated approach for diagnosing conditions like cortical malformation (CM) and periventricular nodular heterotopia (PVNH).
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Neurology
Background:
- Malformations of cortical development (MCD) are a group of congenital neurological disorders.
- Accurate and early detection of MCD is crucial for effective patient management.
- Current diagnostic methods can be time-consuming and may require expert interpretation.
Purpose of the Study:
- To develop and validate a deep learning model for automated detection of MCD.
- To differentiate between specific MCD subtypes, including diffuse cortical malformation (CM) and periventricular nodular heterotopia (PVNH), and normal brain MRI.
- To assess the performance of various convolutional neural network (CNN) architectures in this diagnostic task.
Main Methods:
- Trained four distinct CNN architectures on a dataset of brain MRI images.
- Employed techniques such as batch normalization, global average pooling, dropout, transfer learning, and data augmentation to enhance model robustness and prevent overfitting.
- Utilized a dataset comprising normal MRI scans, CM, and PVNH cases, ensuring no subject overlap between training, validation, and test sets.
Main Results:
- The InceptionResNetV2 CNN architecture demonstrated superior performance across all tested models.
- In the test set, the model achieved an area under the curve (AUC) of 0.89 and accuracy of 0.81 for distinguishing CM from normal MRI.
- For differentiating PVNH from normal MRI, the model achieved an AUC of 0.90 and accuracy of 0.84. The three-class classification (CM, PVNH, normal MRI) yielded an AUC of 0.88 and accuracy of 0.74.
- Saliency maps confirmed that the models focused on diagnostically relevant image regions.
Conclusions:
- CNNs can effectively detect MCD with clinically relevant performance.
- The developed deep learning model offers a fully automated workflow for MCD detection.
- This automated approach eliminates the need for manual image feature selection, potentially streamlining the diagnostic process.
