Evaluation of breast cancer molecular subtypes using artificial intelligence in micro-FTIR hyperspectral images

Abstract

Breast cancer is the most incident cancer worldwide. The evaluation of molecular subtypes and their biomarkers plays an essential role in prognosis. The biomarkers used are Estrogen Receptor (ER), Progesterone Receptor (PR), Human Epidermal growth factor Receptor-type 2 (HER2), and Ki67. Based on these, subtypes are classified as Luminal A (LA), Luminal B (LB), HER2 subtype, and Triple-Negative Breast Cancer (TNBC). The gold standard for this analysis is histology and immunohistochemistry, semi-quantitative techniques that present inter-laboratory and inter-observer variations. The Fourier Transform Infrared micro-spectroscopy (micro-FTIR), which provides hyperspectral images with biochemical information of biological tissues, is applied together with artificial intelligence (AI) for cancer evaluation. In this thesis, twenty samples of two breast cancer cell lines, BT-474 and SK-BR-3, were used to define the optimal number of co-added scans for machine learning (ML) techniques. Linear Discriminant Analysis (LDA), Partial Least Squares Discriminant Analysis (PLS-DA), K-Nearest Neighbors (KNN), Support Vector Machines (SVM), Random Forest (RF), and Extreme Gradient Boosting (XGB) models were used. Sixty hyperspectral images of 320x320 pixels were collected from thirty patients of a human breast biopsies microarray, each containing a breast cancer (CA) and an adjacent tissue (AT) core. Automated methods based on K-Means clustering were developed for data organization and pre-processing to one-dimensional (1D) and two-dimensional (2D) data. The dataset was used to train two new deep learning models for breast cancer subtype evaluation: CaReNet-V1, a 1D Convolutional Neural Network (CNN); and CaReNet-V2, a 2D CNN. All ML models achieved similar performances with the b256_064 (256 background scans and 64 sample scans), b256_128, and b128_128 groups, where the best accuracy of 0.995 was presented by the XGB model. The b256_064 was established as the ideal among the three due to the shortest acquisition time. The K-Means-based method enabled fully automated preprocessing and organization, improving data quality and optimizing CNN training. CaReNet-V1 effectively classified CA and AT (individual spectra test accuracy of 0.89), as well as HER2 and TNBC subtypes (0.83 and 0.86), with greater difficulty for LA and LB (0.74 and 0.68). The model enabled the evaluation of the most contributing wavenumbers to the predictions, providing a direct relationship with the biochemical content of the samples. CaReNet-V2 demonstrated better performance than 1D, with test accuracies above 0.84, and enabled the prediction of ER, PR, and HER2 levels, where borderline values showed lower performance (minimum accuracy of 0.54). The Ki67 percentage regression demonstrated an absolute mean error of 3.6%. On the other hand, its impact evaluation by wavenumber was inferior to 1D. Thus, this study indicates image-based AI techniques using micro-FTIR as potential providers of additional information to pathological reports, also serving as patient biopsy screening techniques.