JMIR Publications

ABSTRACT

Recent developments in deep learning have led to a growing interest in computer-assisted diagnosis and the automated detection of disease in medical images. However, many of these technologies remain experimental and lack widespread use in clinical environments. This is largely because classification decisions of deep learning are often made without concrete explanations that are easily interpretable by clinicians. As such, this study employed gradient-weighted class activation mapping (Grad-CAM) to generate enhanced visualization heatmaps, for improved interpretability of classification The viability of this technique was evaluated using a set of functional magnetic resonance imaging (fMRI) scans acquired from migraine patients. A convolutional neural network (CNN) was used in combination with three fMRI indicators: the amplitude of low-frequency fluctuations (ALFF), regional homogeneity, and regional functional correlation strength, to distinguish two migraine subtypes from healthy controls. These data included scans from 21 migraine patients without aura, 15 patients with aura, and 28 healthy controls. The area under the receiver operating characteristic curve was 0.99 for the test set. Three common CNN architectures (Vgg16, GoogleNet, and ResNet18) were used to train the model, though the network type played only a marginal role in classification accuracy. The selection of indicators was far more effective for improving performance. A series of heatmaps suggested ALFF to be the most informative activation metric. The consistency of activated brain regions in migraine patients confirmed the improved interpretability of classification results available with the proposed technique, which could increase the clinical relevance of deep learning models.