Hyperspectral Band Selection with Learnable Weights for Efficient Glioblastoma Detection

<p dir="ltr">This study explores hyperspectral imaging (HSI) as a non-invasive method for glioblastoma detection, leveraging band selection with learnable weights to improve efficiency. Using the HistologyHSI-GB dataset of 469 images from 13 patients, data was preprocessed through calibration, normalization, band reduction, and patching. A three-layer 2D CNN model was compared against logistic regression and SVM, achieving superior performance with an AUC of 0.9396 and accuracy of 87.74%. Learnable band selection identified key spectral ranges (468–484 nm, 582–630 nm, 677–730 nm), including alignment with protoporphyrin IX (PpIX), a known glioblastoma biomarker. Models trained on reduced bands (16, 32, 56, 112) achieved comparable or better accuracy than the full 275-band input, with the 16-band model reaching 89.9% accuracy. These findings demonstrate that optimized band selection can reduce computational load while maintaining diagnostic accuracy. Future work will refine minimal band sets and co-design imaging hardware tailored for glioblastoma diagnosis, advancing HSI as a practical clinical tool.</p>