Abstract
Breast cancer remains a major global health burden, and microcalcifications are an important feature frequently seen on mammograms. These microcalcifications occur in two distinct chemical forms, calcium oxalate and hydroxyapatite. However, only hydroxyapatite is potentially associated with malignancy. Conventional mammography cannot reliably distinguish between these two types, which motivates the development of complementary diagnostic techniques.
Low wavenumber Raman spectroscopy is investigated in this thesis as a potential adjunct to mammography for microcalcification classification. Its performance is assessed through the classification of synthetic microcalcifications embedded at varying depths within a series of human breast tissue phantoms, including wax, chicken tissue, and hydrogel, with and without simulated skin layers of differing pigmentation. Machine learning models based on support vector machines (SVMs) and convolutional neural networks (CNNs) were developed and evaluated using these Raman spectral datasets. In wax tissue phantoms, three Raman system configurations were compared. The defocused micro–spatially offset low wavenumber Raman spectroscopy (D-microSOLWRS) configuration achieved the highest performance, yielding an area under the curve (AUC) of 0.94 based on receiver operating characteristic (ROC) analysis, demonstrating reliable classification of microcalcifications at varying burial depths. Across two of the three systems, the low wavenumber spectral region provided superior classification performance. Deviations from this trend were attributed to the effects of silica glass subtraction, which disproportionately affected low wavenumber signals.
The robustness of SVM and CNN approaches was further examined using transmission Raman spectra of microcalcifications embedded in chicken tissue with intentionally introduced spectral artefacts. While the SVM model achieved a higher peak ROC AUC of 0.989, it was found to be more sensitive to artefacts than the CNN model. This highlights the tradeoffs between performance and robustness in complex Raman datasets. A novel optical configuration, termed dual beam Raman (DBR), was proposed to simultaneously acquire transmission and backscattered Raman signals, while illuminating the sample from opposing sides. Using sulfur embedded in hydrogel phantoms, Raman signals were detected at depths of approximately 5mm. Additionally, the independent isolation of transmission and backscattered signals was demonstrated. When applied to microcalcification classification in hydrogel phantoms, the DBR system significantly outperformed D-microSOLWRS, achieving an ROC AUC of 1.00 compared to 0.65.
This performance advantage was further evaluated using a more chemically complex dataset comprising carbonate and magnesium substituted hydroxyapatite. The DBR system again outperformed D-microSOLWRS, achieving ROC AUCs of 0.66 and 0.53, respectively. The reduced performance was attributed to the inclusion of two different concentrations of substitutions (2.5 and 5 wt%), which increased spectral overlap between classes. When these concentrations were grouped, classification performance improved to ROC AUCs of 0.86 for DBR and 0.59 for D-microSOLWRS, a trend confirmed by models trained exclusively on Type II spectra.
Finally, a more physiologically realistic hydrogel phantom incorporating fat layers and skin layers with varying melanin content was developed to investigate the influence of skin pigmentation on classification accuracy. Increasing melanin content was found to significantly degrade SVM classification performance, with accuracy decreasing from 0.83 for very light skin tones to 0.33 for dark skin tones. This effect was mitigated by constructing models using selected spectral regions of interest, improving classification accuracy for dark skin tones to 0.72.
Overall, this work demonstrates the potential of low wavenumber Raman spectroscopy, particularly when combined with dual beam acquisition and appropriate machine learning strategies, as a complementary technique for microcalcification classification in breast cancer diagnostics.