Automated detection of polymicrogyria in pediatric patients using deep learning

Polymicrogyria (PMG) is a multifaceted neurological disorder caused by abnormal cortical folding, mostly in children. It commonly results in developmental delays, seizures, and motor weakness. The mild features of PMG in neuroimaging often make its identification difficult, even for experts. In this paper, we assess the efficacy of various advanced image preprocessing strategies on the overall performance of Convolutional Neural Network (CNN) applied for PMG diagnosis in MRI brain scans. We employ a pre-processing sequence that includes Min-Max normalization, Contrast Limited Adaptive Histogram Equalization (CLAHE), Bilateral filtering, and Canny edge detection aimed at improving the recognition of subtle features without losing essential details. The techniques can enhance the visualization of delicate structural deformities in the brain MRI images and assist in the diagnosis of neurological disorders by clinicians. Experimental results suggest that performance enhancement was achieved with all of the tested CNN architectures. ResNet-101 has exhibited the most remarkable accuracy enhancement by 10.3%. ResNet and VGG architectures delivered much greater performance improvement as compared to MobileNetV2 and DenseNet-201 models. GradCAM++ is adopted to infer the decision-making mechanism of the considered deep learning architectures. The methodology finds applications in neurological imaging and may be used to assist healthcare providers in the diagnosis of polymicrogyria. Our findings emphasize the crucial role of image pre-processing techniques in increasing the capabilities of deep learning frameworks to assist with complex tasks in medical image analysis.