A deep learning framework for Segmentation of acute ischemic stroke lesions on multimodal MRI images

Background: Accurate segmentation of stroke lesions on MRI images is very important for neurologistsin planning of post stroke care. Segmentation helps clinicians to better diagnosis andevaluation of the any treatment risks. However, manual segmentation of brain lesions relies onthe experience of neurologists and is also a very tedious and time-consuming process. So, in thisstudy, we proposed a deep convolutional neural network (CNN-Res) that automatically performssegmentation of ischemic stroke lesions from multimodal MRIs.Methods: CNN-Res used a U-shaped structure, so the network has encryption and decryptionpaths. The remaining units are embedded in the encoder path. In this model, to reduce gradientdescent, the remaining units were used and to extract more complex information in images,multimodal MRI data were applied. In the link between the encryption and decryption subnets,the bottleneck strategy was used, which has reduced the number of parameters and training timecompared to similar research.Results: CNN-Res was evaluated on two distinct datasets. First, it was examined on a dataset collectedfrom the Neuroscience Center of Tabriz University of Medical Sciences, where the averageDice coefficient was equal to ۸۵.۴۳%. Then, to compare the efficiency and performance of themodel with other similar works, CNN-Res was evaluated on the popular SPES ۲۰۱۵ competitiondataset where the average Dice coefficient was ۷۹.۲۳%.Conclusion: This study presented a new and accurate method for segmentation of MRI medicalimages using deep convolutional neural networks called CNN-Res, which directly predicts segmentmaps from raw input pixels.

کلیدواژه ها:

ischemic stroke ، Convolutional network ، Lesion segmentation ، MRI ، Informatics ، Deep learning