SpecNet: A hyperspectral band-learning method for slice-level differentiation of fibroadenoma and phyllodes tumor

This retrospective single-center study aimed to explore the discriminative potential of visible-near-infrared (400-1000 nm) hyperspectral imaging (HSI) combined with a lightweight and interpretable deep-learning model for differentiating fibroadenoma (FA) from phyllodes tumor (PT). The approach also provides clinically relevant spectral evidence for pathology.

Formalin-fixed paraffin-embedded(FFPE) sections from 215 patients (105 FA, 110 PT) were retrospectively collected between January 2023 and July 2025. Each one-dimensional reflectance spectrum underwent standardized preprocessing. A compact deep-learning framework (SpecNet) was developed using convolutional and attention modules to capture both local spectral variations and long-range dependencies. Five-fold cross-validation was performed within the dataset. External validation was not included due to the retrospective and exploratory design. For comparison, three baseline classifiers (1D-CNN, Transformer-SVM, and XGBoost) were evaluated. The main metric was AUC, along with Accuracy, Sensitivity, and Specificity.

SpecNet achieved an AUC of 0.89, with accuracy 85.21 %, sensitivity 80.73 %, and specificity 87.45 %. It consistently outperformed baseline models. Spectral analysis identified 540-590 nm and 700-820 nm as key diagnostic regions, corresponding to hemoglobin absorption and near-infrared water scattering.

SpecNet demonstrated robust differentiation between FA and PT using ex-vivo HSI data from a single institution. Although external and prospective validation are needed, the method shows potential for clinical applications such as biopsy triage and intraoperative margin assessment.