Ex Vivo Tissue Classification Using Broadband Hyperspectral Imaging Endoscopy and Artificial Intelligence: A Pilot Study

Endoscopy stands as a prevalent method for cancer detection; however, there are limitations in its diagnostic ability and time efficiency for clinical applications. In this study, we address these limitations by employing hyperspectral imaging (HSI) across visible, near-infrared, and short-wavelength infrared (SWIR) wavelengths. Utilizing machine learning and deep learning frameworks, we develop a diagnostic probe capable of ultrabroadband hyperspectral sensing, subsequently enhancing existing classification techniques. Hyperspectral datacubes from breast tissue samples are captured and serve as the foundation for our models. Machine learning generates optimal models by fitting against multiple possibilities, while deep learning employs an advanced convolutional neural network (CNN). Both methods yield promising real-time classification results, with accuracies reaching up to 90%. While deep learning models exhibit superior accuracy, their training processes require larger datasets and correspondingly lengthened time spans, balancing the overall utility of both frameworks. Our research is a big step forward toward more precise and time-efficient endoscopic cancer sensing and diagnosis.

I. Introduction

Cancer persists as one of the most significant health issues worldwide. It was approximated that in 2020 alone, there were 19.3 million new cancer cases accompanied by 10.0 million deaths globally [1]. As a result, numerous endeavors have been undertaken to enhance early detection and treatment techniques with the aim of reducing the incidence of cancer-related illnesses and deaths.