By Dr. Matteo Pallocca
Precision Medicine is now a common practice and an established diagnostic paradigm, revolutionizing therapeutic approaches thanks to novel multidisciplinary schemes. One pillar of these new techniques is Artificial Intelligence (AI), the umbrella world defining all the computational techniques able to mimic or reproduce intelligent behaviors, such as language, pattern recognition, or classification. These techniques are able to provide novel models of prediction and prognosis to our patients in order to enable a new era of High-Resolution, other than Precision, Medicine.
With the purpose to give a basic framework and round table of discussion to all the several units dealing with structured and unstructured data in the Institute, the Scientific Direction prompted the creation of the Interdepartmental group of Artificial Intelligence and Imaging. The final aim is also to accelerate the Digitalization process of several data types and the employment of complex AI models on biomedical data.
The team is interdisciplinary in nature and comprises Bioinformaticians, Radiologists, Medical Physicists, Engineers, Medical Doctors, Biologists, Physics, Statisticians, and Experts in Scientific Communication.
Regarding unstructured data, Images are the bulkiest and largest set of data generated in a research hospital setting. Radiological data, for instance, are born digital, but they do need to undergo several in silico processing steps in order to extract modeling-ready numeric features. Furthermore, the Pathology Department of a cancer center processes thousands of Immunohistochemistry slides via human analysis and data curation. Now, the automated digitalization of said images enables an unprecedented power to reanalyze with novel techniques and algorithms hundreds of patient slides altogether and to overcome the intrinsic inter-operator human variability.
When it comes to structured data, -omics are the bread and butter of many diagnostic Units, with Next Generation Sequencing being applied to thousands of patients (ref. Genomics group), along with several other facilities such as Lipidomics and Metabolomics. The future of AI in Precision Medicine lies in multi-omics integration, with its numerous technical challenges due to data heterogeneity and batch effect distribution.
Novel synergies among Units
During the first months, the main focus of the AI and Imaging group has been the presentation of several modeling and analysis activities among Units that have been physically and strategically separated, such as Radiology/Medical Physics with the Genomics/Bioinformatics department.
For instance, the Radiology unit and the Medical Physics Laboratory shared their experience on Radiomics analysis of oropharyngeal squamous cell carcinomas and Head and Neck cancer. These projects were strongly related to AI not only for the feature extraction methods but also for the Machine Learning models employed on imaging features that exhibited a classification accuracy of over 90% for benign/malignant parotid lesions. Another joint-venture ongoing with the Galeazzi Orthopaedic Institute is focused on the Texture Analysis of Rare Tumor lesions, with the intent to better separate benign from malignant lesions from the TC and RM data.
Furthermore, during 2020 the interaction between the Digital Pathology and the Bioinformatics group enabled the expansion of the implementation of the Aperio AT2 System for digitalization with the GENIE tool for automatized segmentation of digitalized slides. This tool enables to train the macro-regions of interests of each slide, to then apply the Aperio scoring algorithms to identify immunohistochemical staining specific to each cell in every region (such as tumor, stroma, etc).
This system enabled to accelerate two projects already in place employing Digital Pathology: a digitalization effort on PD-L1 staining on Head and Neck Cancer and another casuistry of Non-Small Cell Lung Cancer treated with Immuno-Checkpoint inhibitors stained with PD-L1 and other CD* immunological markers, with the intent to define a novel predictive immunoscore (ref. Immunology Unit).
The Lung Radiogenomic Pilot
The first project stemmed from the AI & Imaging group pertains a complete multi-omic profiling of a casuistry of 150 Non-Small Cell Lung Cancer Patients who underwent surgery in our Thoracic Surgery department. These patients have pre-surgery TC scans from our Radiology Unit, complete oncogene sequencing from the Pathology, and clinical data annotation concerning treatments, progression events, and comorbidities such as smoking. This multi-omics integrative analysis will enable to shear light on how molecular mechanisms such as somatic mutations influence imaging data and whether a combined radiogenomic model improves biomarker modeling for prognostic and predictive endpoints.
Digitalization and computational image analysis of CD8 staining in Non-Small Cell Lung Cancer in Digital Pathology