Deep learning based automation of mean linear intercept quantification in COPD research

Lars Leyendecker^1*Louisa Weltin²Florian Nienhaus³Michaela Matthey⁴Bastian Nießing²Daniela Wenzel⁴Robert Schmitt⁵

• ¹Department of Production Quality, Fraunhofer Institute for Production Technology (FHG), Aachen, Germany
• ²Department of Bio-Adaptive Production, Fraunhofer Institute for Production Technology (FHG), Aachen, Germany
• ³Department of Production Metrology, Fraunhofer Institute for Production Technology (FHG), Aachen, Germany
• ⁴Department of Systems Physiology, Faculty of Medicine, Ruhr University Bochum, Bochum, North Rhine-Westphalia, Germany
• ⁵Chair of Intelligence in Quality Sensing, Laboratory for Machine Tools and Production Engineering, RWTH Aachen University, Aachen, Germany

Chronic obstructuve pulmonary disease (COPD), a major cause of global mortality, necessitates novel therapies targeting lung function and remodeling. Their effect on emphysema formation is initially investigated using mouse models by analyzing histological lung sections. The extent of airspace enlargement that is characteristic for emphysema is quantified by manual assessment of the mean linear intercept (MLI) across multiple histological microscopy images. Besides being tedious and cost intensive, this manual task lacks scientific comparability due to complexity and subjectivity. In order to continue with the well-established practice and to preserve the comparability of study results, we propose a deep learning-based approach for automating the determination of MLI in histological lung sections utilizing the AutoML software AIxCell which is specialized for the domain of semantic segmentation-based cell culture and tissue analysis.We develop and evaluate our image processing pipeline on stained histological microscope images that stem from a study including two groups of C57BL/6 mice where one group was exposed to cigarette smoke while the control group was not. The results indicate that the AIxCell segmentation algorithm achieves excellent performance, with IoU scores consistently exceeding 90 %. Furthermore, the automated approach consistently yields higher MLI values compared to the manually generated values. However, the consistent nature of this discrepancy suggests that the automated approach can be reliably employed without any limitations. Moreover, it demonstrates statistical significance in distinguishing between smoker's and non-smoker's lungs.