Weakly supervised anomaly detection coupled with Fourier transform infrared (FT-IR) spectroscopy for the identification of non-normal tissue.

The detection and classification of histopathological abnormal tissue constituents using machine learning (ML) techniques generally requires example data for each tissue or cell type of interest. This creates problems for studies on tissue that will have few regions of interest, or for those looking to identify and classify diseases of rarity, resulting in inadequate sample sizes from which to build multivariate and ML models. Regarding the impact on vibrational spectroscopy, specifically infrared (IR) spectroscopy, low numbers of samples may result in ineffective modelling of the chemical composition of sample groups, resulting in detection and classification errors. Anomaly detection may be a solution to this problem, enabling users to effectively model tissue constituents considered to represent normal tissue to capture any abnormal tissue and identify instances of non-normal tissue, be it disease or spectral artefacts. This work illustrates how a novel approach using a weakly supervised anomaly detection algorithm paired with IR microscopy can detect non-normal tissue spectra. In addition to incidental interferents such as hair, dust, and tissue scratches, the algorithm can also detect regions of diseased tissue. The model is never introduced to instances of these groups, training solely on healthy control data using only the IR spectral fingerprint region. This approach is demonstrated using liver tissue data from an agrochemical exposure mouse study.

Infrared micro-spectroscopy coupled with multivariate and machine learning techniques for cancer classification in tissue: a comparison of classification method, performance, and pre-processing technique.

The visual detection, classification, and differentiation of cancers within tissues of clinical patients is an extremely difficult and time-consuming process with severe diagnosis implications. To this end, many computational approaches have been developed to analyse tissue samples to supplement histological cancer diagnoses. One approach is the interrogation of the chemical composition of the actual tissue samples through the utilisation of vibrational spectroscopy, specifically Infrared (IR) spectroscopy. Cancerous tissue can be detected by analysing the molecular vibration patterns of tissues undergoing IR irradiation, and even graded, with multivariate and Machine Learning (ML) techniques. This publication serves to review and highlight the potential for the application of infrared microscopy techniques such as Fourier Transform Infrared Spectroscopy (FTIR) and Quantum Cascade Laser Infrared Spectroscopy (QCL), as a means to improve diagnostic accuracy and allow earlier detection of human neoplastic disease. This review provides an overview of the detection and classification of different cancerous tissues using FTIR spectroscopy paired with multivariate and ML techniques, using the F1-Score as a quantitative metric for direct comparison of model performances. Comparisons also extend to data handling techniques, with a provision of a suggested pre-processing protocol for future studies alongside suggestions as to reporting standards for future publication.