Student Name: Rigon Sallauka
Summary:
I am Rigon Sallauka, Doctoral Candidate no. 8 in BosomShield project. My upbringing took place in Prizren, Kosovo, where I completed both my primary and secondary education. In 2014, I embarked on my higher education journey in Tirana, Albania, culminating in my graduation in 2019. At the Faculty of Natural Sciences, University of Tirana, I earned both a Bachelor's and a Master's degree in Mathematics. It was in Tirana where I first started to work in technology. My first job was in a development company. Following this, I returned to my hometown, where I started to work in academia. For approximately three years, I served as a teaching assistant at the Faculty of Computer Science, University of Prizren, Kosovo, before transitioning to my current involvement in this esteemed project. Presently, I am doing my PhD studies in the Faculty of Electrical Engineering and Computer Science, University of Maribor, Slovenia, where I also work in the position of Researcher.
Objectives:The main objective is to build a prediction model combining three types of data; histopathological, radiological biomarkers and clinical readouts of patient’s EHR for developing BC relapse prediction. To this end, multi-modal biomarkers extracted by pathologists, radiologists, and oncologists from both WSI and radiology modalities along with clinical data will be exploited for estimating the risk of BC relapse. Several stages are identified prior to the modelling phase: 1) different trustworthy classifiers for relapse prediction will be studied, 2) relevant biomarkers analysed by medical experts will be identified and considered for modelling, 3) studying the similarity and dissimilarity within the three types of data, 4) analysing the existing correlations between the different input modalities that will simultaneously allow the risk estimation of BC relapse, and finally, 5) constructing an explainable and interpretable prediction model for estimating the risk of BC relapse based on a) radiological biomarkers b) combination radiological and histopathological biomarkers, and c) a combination of radiological biomarkers, histopathological biomarkers and clinical data. This project will include a comparison between CNNs without prior feature extraction and others such as daboost and Random Forest with the features extracted as a pre-processing stage. Different architectures of CNNs and RNNs will be implemented to combine the three data in one efficient model. The final obtained model will be validated in a population of patients to assess how it fits in the new cohort.
Outcomes:
The expected outcomes are: 1) identification of most relevant biomarkers between histopathological, radiological and clinical data. 2) The discovered meaningful properties of the different pathologic image biomarkers and radiologic-image features associated to them. 3) Comparison of the state-of-the-art BC relapse prediction models. 4) BC relapse predictor based on only radiological biomarkers. 5) a trustworthy AI-based model enabling the explainability and combing three types of biomarkers for accurate relapse risk estimation.