FS-FCN: a frequency-spatial fusion convolutional network for brain tumor classification in MR images

Brain tumor classification using magnetic resonance imaging (MRI) is important for computer-aided medical image analysis. However, different brain tumor types may exhibit highly similar grayscale patterns, textures, and boundaries. Conventional convolutional neural networks rely primarily on spatial-domain convolution and do not explicitly model frequency-domain information. Moreover, high-frequency details, such as edges and textures, are often weakened by multi-stage downsampling, while interaction between spatial- and frequency-domain features remains limited. Consequently, fine-grained differences among tumor categories may not be effectively captured.

To address these limitations, a Frequency-Spatial Fusion Convolutional Network (FS-FCN) was developed for brain tumor MRI classification. A Parametric Wavelet Downsampling (PWD) module and a Frequency-Spatial Convolution (FSConv) module were incorporated into the convolutional network. Spatial- and frequency-domain information was jointly modeled during multilevel encoding and deep feature learning to improve the preservation and discriminative representation of key structures, local textures, and edge details.

On the Cheng Brain Tumor MRI Three-Class Classification Dataset, FS-FCN achieved an average classification accuracy of 98.53% ± 0.35% using five-fold cross-validation. The applicability of the model across different classification settings was further evaluated through a four-class classification experiment.

The results demonstrate that jointly modeling spatial- and frequency-domain information improves the representation of brain tumor MRI features and provides an effective feature-learning approach for brain tumor MRI classification.