Breast cancer is a heterogeneous disease comprising distinct molecular subtypes that require accurate diagnosis for effective treatment. Conventional diagnostic methods primarily rely on histologic evaluation, which limits access to molecular-level information. Here, we present a label-free strategy combining Fourier-transform infrared (FTIR) imaging with machine learning for breast cancer diagnosis and subtype classification from tissue sections. A two-step approach was developed: first, to discriminate benign from malignant breast tissues, and second, to differentiate HER2+ from triple negative breast cancer (TNBC) subtypes. Using a cohort of 61 patients, LightGBM-based models reached sensitivities above 84% in distinguishing benign from malignant lesions, with ensemble learning achieving 100% accuracy across independent test sets, and 87.5% when including luminal tumors not used during training. Subtype classification proved more challenging but integrating the tumor microenvironment into the models improved performance, reaching sensitivities of 85.7% for HER2+ and 83.3% for TNBC patients. This study demonstrates, for the first time, the potential of FTIR imaging combined with AI to distinguish between HER2+ and TNBC subtypes directly from tissue sections. These findings highlight the relevance of spectral histopathology for rapid, nondestructive, and accurate breast cancer diagnosis, paving the way for integration into clinical workflows.
Tahech et al. (Tue,) studied this question.