Deep Learning-Based Quantitative Analysis of Cerebral Microbleeds: Segmentation and Classification Using CNN

Cerebral Microbleeds (CMBs) are among the significant contributors to mortality worldwide and require accurate diagnosis for effective medical intervention. Owing to their wide variability in size, shape, and intensity, manual identification and classification of CMBs in brain imaging remain a complex and error-prone task. This study proposes an automated classification framework for brain MRI-filtered images, categorizing them as either normal or abnormal. The suggested methodology combines a tailored Convolutional Neural Network founded on the ResNet50 architecture, employing a blend of image processing and deep learning strategies. First, a number of preprocessing processes were implemented to increase the MRI pictures quality. One of these steps was the fusion of multi-focus images, which helped to make details more visible. These enhanced images were then processed through a 13-layer CNN architecture specifically designed for effective CMB classification. The strength of the proposed CNN-ResNet50 model was confirmed through validation with two independent datasets. Experiment one used a 10-fold cross-validation procedure, while experiment two split the dataset in half, with 80% used for training and 20% for testing. The model achieved a train-test split accuracy of 98.77% and a cross-validation accuracy of 98.33% while classifying Dataset 1. An accuracy of 92.22% and an accuracy of 93.33% were attained by the model in the two experimental setups for Dataset 2. All investigations used real-world MRI scans. This data set originated from Neyyoor, India's CSI Medical Mission Hospital's International Cancer Center (ICC). The efficacy of the suggested CNN-ResNet50 model was evaluated in comparison to established deep learning models, such as AlexNet and the original ResNet50. Experimental data indicate that our proposed method surpasses both comparative models regarding classification accuracy.