Deep Learning Applications in Pre-Clinical Imaging for accelerated Drug Discovery Studies

In development of new drugs, preclinical studies based on animal models are essential to understand the pathogenesis and assess the efficacy of the identified new drug. As technological advancements continue to unfold, medical imaging is gaining significant prominence in the preclinical domain as well. Having already shown state-of-art performaces in clinical imaging, automated tools capable of quickly and effectively processing images can offer invaluable assistance in preclinical imaging endeavors. In this thesis, deep learning (DL) and classical image processing methodologies were designed and developed to support pharmacologists during drug discovery studies for idiophathic pulmonary fibrosis (IPF) in both in-vivo (Micro-CT imaging) and ex-vivo (histological imaging) analysis.Particularly in the context of drug research, the integration of heterogeneous data from different sources is of particular interest, as it allows phenomena to be analyzed from multiple perspectives. For this reason, this thesis focuses on the automated analysis of different imaging modalities (micro-ct and histology) and their integration. In Micro-CT imaging, deep learning is exploited to perform automatic segmentation of lungs and airways. Lung segmentation is then used to extract lung densitometry by exploiting the linear relationship between X-ray attenuation and tissue density, while airways segmentation is used to extract geometric measurements, facilitating the assessment of morphological changes in the airways at various stages of IPF development. In histological imaging, deep learning is used to create a local map by assigning a score to each patch of the created tesselation, representing the severity of fibrosis. Finally, since histology involves less than $1\%$ of lung volume, a semi-automated approach to identify the micro-CT slice corresponding to the histologic slice was implemented, enabling comparisons between in-vivo and ex-vivo evaluations. For Micro-CT images, the developed pipeline provided better results compared to state-of-the-art approaches. Lung densitometry showed results comparable to manual evaluations, and the geometric assessment of airways yielded promising results. Regarding histological images, the deep learning model achieved a level of accuracy similar to that obtained through manual assessments in assigning the degree of fibrosis to different areas of lung tissue. The matching of the two imaging modalities showed promising results with a level of accuracy comparable to human error. All three tools significantly reduced the time required for manual analyses and proved to be effective enough to be integrated into the routine work of pharmacologists.\\ This work presents a platform for automated analysis of micro-CT scans and histological images of mouse models of idiopathic pulmonary fibrosis. This automated tool enables rapid and comprehensive data analysis, thus laying the foundation for its wider use in preclinical settings.