Convolutional-morphological neural network applied to medical imaging

Abstract

Mathematical morphology, a versatile technique frequently employed in image processing, finds wide-ranging applications in tasks such as segmentation, filtering, compression, edge detection, and feature extraction. In this study, we focus on the latter application and evaluate the effectiveness of combining mathematical morphology operations with convolution in a deep learning framework, comparing it with handcrafted features paired with traditional machine learning classifiers. To this end, we introduce the Morphological and Convolutional Neural Network (MCNN) trained with Extreme Learning Machines (ELM). Our methodology encompasses an internal assessment, a comparison with three commonly referenced Convolutional Neural Networks (CNNs) - ResNet-18, ShuffleNet-V2, and MobileNet-V2, and an evaluation of performance across four distinct optimizers - ELM, SGD, ADAM, and RProp. We conduct four classification tasks using both the MCNN approach and classic machine learning techniques for glaucoma, melanoma, breast cancer, and COVID-19 detection, leveraging the ORIGA, ISIC, RSNA, and Thermal-Covid datasets, respectively. While the classification performance for glaucoma and melanoma proved reliable, with accuracies of 0.73 (0.67, 0.80) 95% CI and 0.88 (0.84, 0.92) 95% CI, breast cancer and COVID-19 exhibited random classification, yielding accuracies of 0.50 (0.45, 0.57) 95% CI and 0.52 (0.47, 0.57) 95% CI. In this work, we offer several contributions: a deeper exploration of mathematical morphology as a feature extractor for medical diagnosis using deep learning, the introduction of the MCNN method incorporating these operations, an analysis of its strengths and weaknesses, a comparative assessment against conventional handcrafted features, and an examination of performance variations with different optimizers when applying morphological operations.