IbnSinaNet: a hybrid CSPDenseNet and squeeze attention under the supervision of group CNNs for automatic segmentation of MS lesions in MRI

Abstract Diagnosing Multiple Sclerosis (MS) relies on identifying brain plaques via Magnetic Resonance Imaging (MRI) and observing clinical symptoms. Given the morphological similarity between the MS lesions and normal brain tissue, as well as the limitations of the naked eye to detect subtle changes in plaque volume or number over time, the implementation of advanced, intelligent detection systems is essential. This study proposes IbnSinaNet, a deep learning framework for automatic MS lesion segmentation. The proposed network consists of three main blocks, including the Transition Down block, DenseNet, and Squeeze Attention under the supervision of the Group Convolutional Neural Network (G-CNNs). The Transition Down block is used for the augmentation of abstract features through the extraction of high-level features. The DenseNet Block is used with Cross Stage Partial Network (CSPNet) in the encoder and without CSPNet in the decoder. Incorporating CSPNet into the encoder significantly reduces the computational cost. Additionally, G-CNNs were employed to enhance the processing speed by considering the independence of input images. We designed the Squeeze Attention block to enhance pixel aggregation, thereby improving dense predictions and pixel-group attention. This led to the establishment of a novel architectural framework, termed IbnSinaNet. We used the widely recognized ISBI 2015 dataset acquired from MS patients to validate the proposed method. Our proposed method has achieved a Dice similarity coefficient of 86.93%, reflecting an improvement of 2.93% over the state-of-the-art MS diagnosis methods. These findings highlight the potential of IbnSinaNet as a robust and precise solution for automatic MS lesion segmentation in MRI, with promising implications for advancing clinical decision support and enhancing patient care. The code is available at https://github.com/MahshidDehghanpour/IbnSinaNet .