The high reoperation rate after breast-conserving surgery (in average 19% in the UK) is associated with the lack of efficient and easy to apply intraoperative methods for detection the tumour residue (“positive margin”) of the excised sample. In-situ tests, based on diffuse reflectance and intrinsic fluorescence spectroscopy could potentially palliate this problem by interrogating tissues at a depth of up to several millimetres. We evaluated three machine learning algorithms applied to a dataset of diffuse reflectance and fluorescence spectra consisting of 181 frozen breast samples, collected from 138 patients. The diagnostic accuracy depended on the applied algorithm and the AUCs ranged from 0.71 to 0.81 (maximal sensitivity 86.16%, specificity 58.97%) and is comparable with existent intraoperational modalities, such as, for example, MarginProbe. Further research is needed to find an optimal combination of spectral features and diagnostic algorithm.
|