Thesis supervisor: Pál Kaposi Novák
Location of studies (in Hungarian): SE Abbreviation of location of studies: SE
Description of the research topic:
Radiomics analysis in abdominal diseases
Ph.D. thesis advisor: dr. Pál Novák Kaposi, assistant professor, Department of Radiology, Medical Imaging Center, Semmelweis University
Hypothesis
Cross-sectional imaging studies are essential for early diagnosis and classification of both benign and malignant tumors in abdominal organs. Radiology reporting conventionally uses semantic image features (i.e., location, shape, contours) described by the radiologist to categorize lesions. The diagnostic criteria used in everyday practice only include a few necessary measurements such as size or attenuation. Meanwhile, tumor heterogeneity assessment with semantic features has remained mainly qualitative and relies heavily on the readers' training and experience. Radiomics is a collective term of different techniques such as texture analysis, which can extract a large number of quantitative features to describe complex distributions of image elements. Texture features describe spatial inter-relationships of pixel grey-level values. The non-invasive characterization of lesions with radiomics features is often more reproducible and better correlate with histopathology diagnosis than conventional semantic parameters. Some texture parameters have shown a close association with molecular subtypes, prognosis, and therapy response in tumors of various origins. Machine learning and deep learning algorithms are highly suitable for processing a large number of quantitative variables. The feature-based prediction models can be built and cross-validated with AI algorithms. Also, lesion-specific deep texture parameters can be extracted from the filters of a convolutional neural network (CNN), which has been pre-trained for image classification on large image depositories (i.e., ResNet, VggNet). The deep image features can also be used in AI classification models.
Aims
Our research aims to demonstrate that new image analysis tools such as texture analysis can facilitate the diagnosis as well as the classification of tumors in abdominal organs. Therefore our research proposal has multiple aims:
- First, we intend to build an image repository through a retrospective collection of diagnostic imaging scans, including multiple modalities, where a histopathology diagnosis is also available.
- Second, we complete the systemic extraction of image features by methods of both texture analysis and convoluted neural network-based extraction of deep image features to establish a link between characteristic image biomarkers and unique phenotypes.
- Third, we will construct artificial intelligence (AI) based models, which are more reliable in the classification of tumors compared to classical radiology diagnostic signs. We will also test the accuracy of the proposed models by extensive cross-validation and utilizing independent validation sets.
- Fourth, we would like to demonstrate that image features can be correlated with clinical, histopathology, and genetic variables to improve prognostication or therapeutic decision making. Moreover, we hypothesize that an algorithmic evaluation of lesion heterogeneity may reveal underlying pathologies, which remain undetected with conventional imaging.
- Fifth, we intend to set up a multi-center collaboration to demonstrate that newly identified radiomics features are reproducible across various scanning conditions, and can be quickly implemented for routine diagnostics of abdominal lesions.
Expected results
By the time all steps of the radiomics pipeline get developed, it will significantly transform the diagnostic workflow of abdominal radiology studies, and facilitate the more precise evaluation of abdominal tumors. An annotated image repository will be developed as a part of the project, which can be the foundation of multiple analyses aimed to answer various classification problems. We will build classification models to diagnose different histology types of liver, pancreas, kidney, and intestinal tumors with better accuracy than conventional radiology evaluation. Accurate non-invasive characterization of tumor types can promote early detection of malignancies and in selected cases, can even substitute for invasive tests such as biopsy. We will also develop prediction models to correlate different tumor phenotypes identified through radiomics with differences in tumor biology such as response to treatment, recurrence or survival. Previous radiomics studies have already shown initial success in some of the tumor classification tasks. Our group, for example, was able to differentiate between various stages of liver fibrosis using texture analysis and machine learning models. Nevertheless, the comprehensive evaluation of the radiomics features in abdominal tumors has not been completed, especially as both feature extraction and AI techniques are still rapidly evolving. In the second stage of the project, we will establish collaboration with other imaging centers in order to complete a multi-center validation of some of the highly useful radiomics features used for tumor classification and prognostication. We will validate the features extensively in various technical settings to put our models to work in everyday practice. Our research proposal has distinct strengths in multiple areas. Through our host institution, which is a leading center in abdominal imaging in our country, we access to a great diversity of cases and employ the most advanced imaging techniques. Our collaborators have a wide range of expertise in diagnostic radiology, image analysis, software engineering, and clinical research. We have already developed and successfully tested elements of the research proposal in prior studies.
Login
