Optical Spectroscopy and Machine Learning for Tumour Classification In Heterogeneous Breast Tissue
Abstract
Purpose
Re-excision rates remain high for early-stage breast cancer patients due to challenges in margin delineation during surgery. Our group has developed a time-resolved fluorescence and diffuse reflectance (TRF-DR) spectroscopy system to address this concern for tumour boundaries where tissue composition is heterogeneous.
Methods
A total of 4818 measurements from 73 frozen patient ex-vivo samples were evaluated. TRF-DR optical features (n=13) were transformed using base-10 logarithm and underwent principal component analysis for feature reduction. A weighted logistic regressor was trained with principal components and pathologist annotations in a patient leave-one-out cross-validation scheme. Tissue measurement areas (1 mm x 1 mm) were pathologist-assigned percentages of low (<25%), medium (25-<75%), and high (≥75%) for tumour, fibroglandular, or adipose composition. Model predictions were assessed via 2 x 2 confusion matrix to calculate the sensitivity and specificity. This was done on the total dataset and on 4 subsets: heterogeneous-only, low, medium, and high tumour amounts.
Results
The model trained with four principal components achieved the following sensitivities and specificities: total – 77%, 68%; heterogeneous-only – 71%, 64%; low tumour – 74%, 67%; medium tumour – 72%, 64%; high tumour – 81%, 70%. For the total dataset, false positive and false negative rates of 32% and 23% were achieved, respectively. Subset analyses demonstrated the best performance for identifying tumour in high amounts and improved classification when including homogeneous tissue compositions during training.
Conclusion
These results show the potential of TRF-DR spectroscopy as a complementary tool for direct-sensing of surgical breast margins. It achieved performance comparable to current clinical modalities (e.g., specimen radiography) under heterogeneous tissue conditions. Future work will involve investigating alternative features and machine learning models that can aid in improving tumour sensitivity in low and medium amounts.