Breast tumor classification based on fluorescence hyperspectral imaging

Rapid frozen sectioning is a commonly used method for obtaining intraoperative pathological diagnosis during breast surgery due to its short turnaround time. This technique enables quick identification of lesion type and pathological grading. However, the process involves several steps including frozen sectioning, staining, and inspection, taking up to an hour to obtain a pathological report. This can still cause significant delays for doctors and patients during surgery, impeding timely diagnosis and intraoperative decision-making. Hyperspectral data can provide valuable auxiliary information for medical pathological diagnosis and enable detection of finer spectral features in histopathology. In this study, fluorescence hyperspectral imaging was utilized to analyze HE-stained and unstained pathological sections of breast tumor tissue. The collected hyperspectral fluorescence data underwent preprocessing operations, such as derivative spectra and data standardization, to minimize the impact of environmental and system noise and enhance the signal-to-noise ratio of hyperspectral data. Subsequently, a three-dimensional convolutional neural network (3D-CNN) classification method that combines spatial structure and spectral information was proposed. The method achieved classification accuracies of 98% and 95% for HE-stained and unstained sections, respectively. These findings demonstrate the potential of using unstained tissue sections for pathological auxiliary diagnosis and provide theoretical support for the use of hyperspectral imaging technology in rapid intraoperative diagnosis of breast tumors.