Distinguishing Low-Grade Chondrosarcoma and Osteochondroma Using Visible-Near Infrared Hyperspectral Spectral Characteristics.

PURPOSE: To evaluate visible-near-infrared hyperspectral imaging (400-1000 nm) combined with a lightweight deep-learning network for differentiating low-grade chondrosarcoma (LGC) from osteochondroma (OC).

METHODS

Spectra were reflectance-calibrated, Savitzky-Golay smoothed, and cropped to 420-850 nm. Multi-level features-single-band reflectance, key band ratios, and PCA components from biologically informative regions were extracted. We developed ChondroSpecNet, a 1D-CNN coupling multi-scale convolutions with a residual squeeze-and-excitation block for end-to-end classification.

RESULTS

ChondroSpecNet yielded AUC 0.92 with 87% accuracy in training, and AUC 0.83 with 86% accuracy on an independent test set. Contribution analysis identified visible-band troughs-especially 472, 467, and 501 nm-as most discriminative by absolute differences and normalized ratios.

CONCLUSION

Hyperspectral microscopy plus a compact network enables accurate, efficient LGC-OC discrimination, offering real-time deployability with robust performance and practical scalability for cartilaginous-tumor diagnosis in clinical workflows.