Benchmarking deep learning architectures for hyperspectral in-vivo brain tumor segmentation

Deep Learning (DL) has become increasingly relevant in medical image analysis, especially for tasks such as surgical guidance. In this respect, Hyperspectral Imaging (HSI) has shown encouraging results for in-vivo brain tumor segmentation during surgeries, where DL models can benefit from spectral information. The literature includes a large collection of Neural Network (NN) architectures designed to process spatial and spectral information in HS images. However, most of these architectures have been developed outside the healthcare domain, and comparative studies on the main architectures applied to HS medical datasets are still limited.

In this work, we provide a systematic benchmark of the principal architectures that have emerged in the last five years, using the 25-band SLIMBRAIN and 128-band HELICoiD datasets. We have also established a taxonomy based on the mechanisms used to build the DL architectures, such as convolutions, attention, and a combination of both.

Some of the convolutional models emerged as the best performing approaches, reaching a maximum F1-score of 65.08% and 91.13% on the two datasets, respectively, followed by hybrid and attention-based ones. Several compact architectures, such as DBDA and RSSAN demonstrated favorable accuracy-efficiency characteristics, showing that high performance is achievable across different spectral dimensionalities without large models.

Beyond providing a comparative ranking, this study identifies architectural mechanisms that are particularly relevant for medical HSI segmentation, including the role of Depth-Wise (DW) convolutions and spatial-spectral feature separation. These findings provide practical design indications for future HSI-based applications in neurosurgery and other medical domains.