SMC-CNN: Stacked Multi-Channel Convolution Neural Network for Predicting Acute Brain Infarct From Magnetic Resonance Imaging Sequences

Acute brain infarct is a major cause of stroke and the second most common cause of fatality worldwide. It’s characterized by abrupt symptoms persisting over 24 hours or leading to death due to blood vessel blockage. There is a need for a fast and automated way to diagnose and predict the outcome of this condition. Medical image analysis has witnessed promising outcomes with the application of deep learning (DL) techniques. To address this problem, we propose two Stacked Multi-Channel Convolutional Neural Networks (SMC-CNNs) for predicting acute infarct using individual and multiple Magnetic Resonance Imaging (MRI) sequences, including Diffusion-Weighted Imaging (DWI), Apparent Diffusion Coefficient (ADC), T2-weighted Fluid-Attenuated Inversion Recovery (T2-FLAIR), and Susceptibility-Weighted Imaging (SWI). We collected de-identified and de-linked MRI sequences from KMC Hospital (Mangalore, India) and compared the efficacy of our models with eight baseline state-of-the-art DL models in an extensive benchmarking study. The collected dataset was pre-processed using a proposed contour-based brain segmentation technique to isolate brain contours from the MRI sequences. These contours were ingested into the two proposed models: Stacked Multi-Channel Convolutional Neural Network for Individual sequences (SMC-CNN-I) and Stacked Multi-Channel Convolutional Neural Network for Multiple sequences (SMC-CNN-M), to predict acute infarct. We conducted experimental evaluations on individual MRI sequences to assess the effectiveness of the models for each sequence and found that the DWI and T2-FLAIR imaging sequences contained more discriminative features for acute infarct prediction than the other sequences. We performed an ablation study by varying and fusing different MRI sequences and observed that the proposed model achieved superior results when all four MRI sequences were used as inputs. We also tested the proposed models on synthetic MRI data generated using a Deep Convolutional Generative Adversarial Network (DCGAN) architecture. We found that the models produced improved results, demonstrating their ability to perform well on real-world data. We conducted a quantitative analysis followed by a qualitative analysis by visualizing infarcts using the Gradient-weighted Class Activation Mapping (Grad-CAM) technique. This demonstrated the model’s ability to detect the precise location of abnormalities.