Advanced U-net segmentation model using residual grouped convolution and attention mechanism for brain tumor MRI image segmentation

The brain is responsible for the "general command" of human thinking and coordination of the body. Thus, various brain diseases can cause great damage to the human body and nervous system. Brain tumors are caused by abnormal cells that grow and multiply irregularly within the brain. Glioma is one of the most common malignant tumors in adults. It originates from glial cells and the surrounding infiltrating tissue, compresses other normal tissues in the brain during the growth process, and blurs the boundary of the tumor. Hence, it is likely to cause damage to normal brain tissue during treatment, causing irreversible harm to human health. Therefore, the precise determination of segmentation boundaries of brain tumors has become an urgent problem to be solved.

With the rapid advancement of medical imaging diagnostic technology, various medical imaging techniques have played important roles in the clinical detection of the disease and formulation of patient treatment plans. Among them, magnetic resonance imaging (MRI) plays an important role in the diagnosis of brain tumors as a noninvasive and safe modality, providing clear and strong contrast. However, the invasive and highly heterogeneous nature of the tumor, leads to a high degree of nonuniformity and borderless features in MRI images. In addition, the influence of external factors and different MRI acquisition methods may change the appearance of tumors or even produce artifacts, reducing image quality. This makes the detection and treatment of brain tumors challenging.

In clinical practice, the labeling of brain tumor target areas is primarily performed by radiologists. Depending on the personal experience of the tagger, this process is not only time-consuming and laborious, but also difficult, as there is no consistent standard for segmentation. Therefore, an accurate and rapid automatic segmentation method would be of great significance to the MR community in academic research and the clinical setting, for the diagnosis, monitoring, and progress evaluation of conditions related to the nervous system, such as brain diseases.

Although deep learning methods such as the convolutional neural network (CNN), can be used for brain tumor image segmentation, the segmentation results based only on the traditional CNN method exhibit rough boundaries and poor performance in tumor details. To solve these problems, this study adopts a residual grouped convolution module (RGCM) based on U-net, to reduce model parameters and computational load and accelerate model convergence. Feature extraction ability and segmentation accuracy of the model are improved, using a convolutional block attention module (CBAM) and a bilinear interpolation method to effectively solve the problems of missing details and unsmooth boundaries in the output segmentation results. Part of this work was presented at the 2022 5th International Conference on Computer Science and Software Engineering under the heading of "RGA-Unet: An improved U-net segmentation model based on residual grouped convolution and convolutional block attention module for brain tumor MRI image segmentation"

The difference between this study and previous work is that, this study investigates the influence of normalization processing, loss function, and network depth on segmentation performance, and further improves the segmentation accuracy of the model. The specific research can be summarized as follows:

(1)

RGCM is used to perform the high-dimensional convolution in the original network layer as eight identical low-dimensional convolutions, thereby reducing the number of parameters, computation load, and convergence time of the model, to obtain higher segmentation accuracy.

(2)

CBAM is used to imitate human visual processing such that the model pays more attention to the feature extraction of the tumor region, increasing the weight of feature learning in that region, to further improve segmentation accuracy.

(3)

The influences of the sampling method, normalization, loss function, network depth, and other factors on segmentation performance were studied, whereby bilinear interpolation upsampling, instance normalization, binary cross entropy (BCE)-Dice loss function, and a four-layer encoder-decoder structure were adopted to improve the proposed model. The proposed model achieved the best segmentation effect.