Reducing Positional Uncertainty in FTIR-Tomography Data Acquisition with Automation and Machine Learning

Fourier Transform Infrared Micro-Spectroscopy (FTIR-MS) is an advanced analytical method employed to examine materials\u27 molecular structure and chemical composition at a microscopic level. By merging the principles of infrared spectroscopy and microscopy, it offers comprehensive insights into the molecular vibrations within a sample. The key components are: (i) infrared spectroscopy which involves analyzing the interaction of infrared light with molecules, (ii) Fourier Transform allows the conversion of a time-domain signal into its frequency-domain representation. In FTIR, an interferogram is obtained, which contains information about the different frequencies present. In summary, FTIR-MS considers infrared spectroscopy to enable the analysis of molecular vibrations and chemical composition at a microscopic level. Whereas tomography is a method of imaging that generates precise, cross-sectional images or slices of an object or sample. By reconstructing collected data, it enables the visualization of internal structures and characteristics of the object. During tomography, the object is measured or imaged from multiple angles, capturing various projections. This study focuses on developing the capability to accurately capture imaging data with FTIR-MS meant for downstream 3D reconstruction with CT. The focus of this study is on measurement of the data and not on the CT-based reconstruction of the acquired FTIR-MS data. The aim is to measure dataset for downstream CT or Limited Angle Computed Tomography (LACT) using FTIR-MS by correcting the positioning problem primarily in the following three steps – (i) lab automation using real-time positioning correction method, (ii) using feature engineering through the use of Principal Component Analysis (PCA) to find the “edge” of the sample and the background, and (iii) microloop’s center detection using two methodologies. A Heuristic Solution is adopted to detect the center by exploiting the microloop’s geometry and the other solution involved image segmentation on synthetic data using Deep Learning (DL) based solution to identify the microloop’s center. To approach a DL based solution, synthetic dataset was generated using variations in microloop structure, sample location and sample shape/size. The UNet based DL model was only trained on synthetic data and had no exposure to the original IR Tomo Data measured in the lab