Vis-SWNIR spectroscopic and hyperspectral imaging sensor integrated with artificial intelligence for early diagnosis of breast cancer

Breast cancer (BC) is a common cancer worldwide, requiring the development of rapid methods for early detection. Recently, advances in artificial intelligence (AI), especially in machine learning (ML), have facilitated the use of hyperspectral imaging (HSI) and portable spectroscopic sensors in disease diagnosis. This paper investigates the use of portable visible-short wavelength near-infrared (Vis-SWNIR) spectroscopy and HSI in the wavelength range of 400–1000 nm to analyze 143 dried plasma spot (DPS) samples, (73 controls and 70 samples from BC patients), using ML for BC diagnosis. In this study, plasma samples were dried and the variability between samples, drying method and factors affecting it were investigated using analysis of variance-simultaneous component analysis (ASCA). The Vis-SWNIR spectroscopic and HSI sensors offer a safe, rapid and cost-effective diagnostic method that is ideal for repeated screening. Due to the complexity of HSI data, multivariate curve resolution-alternating least squares (MCR-ALS) algorithm was used as a feature extraction technique to extract pure spatial and spectral profiles of the existing components. Then, multivariate classification was performed on spectroscopic and HSI data using data driven-soft independent modeling of class analogy (DD-SIMCA), partial least squares-discriminant analysis (PLS-DA), artificial neural networks (ANN), k-nearest neighbor (kNN), random forest (RF) and support vector machine (SVM). The ANN achieved an accuracy 86.0 % in differentiating healthy and diseased samples in HSI data. In contrast, SVM modeling for portable spectrometer data showed an accuracy of 62 % for prediction set. The results showed changes in bilirubin, hemoglobin, porphyrin, proteins and lipids. While the findings for BC detection are promising, more studies are needed. • Portable Vis-NIR and HSI sensors used for breast cancer sample analysis. • MCR-ALS showed a great performance for feature extraction of HSI data. • ANN model reached 85–88 % sensitivity in classifying Vis-NIR HSI plasma data. • SVM on portable Vis-NIR spectrometer gave lower sensitivity (67–70 %). • Key spectral changes linked to bilirubin, hemoglobin, porphyrin, proteins, lipids.