Eduardo Almeda

BASIC INFORMATION

Ph.D. Student: Eduardo Almeda Luna
Advisors: Sebastián Ventura, José María Luna
Started on: October 2021
Keywords: Radiomics, Machine Learning

THESIS PROPOSAL

Radiomics is an emerging term that is becoming increasingly popular in radiology as a diagnostic method for cancer detection. It can be broadly defined as the extraction of hundreds or thousands of quantitative features from medical images, taking advantage of advances in Artificial Intelligence (AI) to provide better performance in classical imaging tests and decision making. Radiomics-based studies follow two clearly differentiated methodologies: classical and deep learning (DL)-based. Focusing on the classical methodology, it is carried out in five different steps: (1) Acquisition of medical images, (2) Image pre-processing, (3) Data segmentation, (4) Feature extraction and (5) Application of AI methods for classification or prediction. In contrast, the DL-based methodology brings together all phases of segmentation, feature extraction and classification in a single step. Despite the rapid advances in radiomics, mostly driven by the great potential of extracting quantitative features from medical images, it is an emerging area open to multiple improvements from both clinical and AI perspectives. In this sense, we consider as a starting hypothesis that the two existing methodologies should not exist without each other, and that their combination can provide a synergy of enormous utility for radiologists, understanding which type of features are usually more related to better diagnoses, which features are generally residual in decision making, or which set of features are highly related and/or have a high discriminant power. We believe that the techniques studied and developed during this PhD thesis can be extremely useful to bring AI and the results extracted from it closer to the end user.

It is noteworthy that radiomics generates hundreds or thousands of features from medical images, which entails an enormous complexity in the estimation of relevant features and their correlations (fat-short problems), increasing the complexity of classification or prediction tasks. Moreover, when we deal with the DL-based methodology, there is a huge problem concerning its zero degree of explainability. A major challenge for radiomics is therefore to combine the goodness of the two methodologies (classical and DL-based) into a single methodology that reduces errors, but also entails a high degree of interpretability and understanding of the models and predictions, as well as knowledge of what might perturb the model. Another challenge of radiomics is to be able to carry out a descriptive analysis of what is happening in the data, both from the point of view of relationships between variables and from the point of view of extracting features that discriminate the data very well.

The overall objective of this research project is to address the above challenges of radiomics by improving the performance of the methodology, increasing the interpretability of the solutions, and applying descriptive analysis techniques on the features extracted by radiomics to improve the understanding of the images. The main objective set out above can be broken down into the following secondary objectives:

• Development of automatic feature extraction and selection methods in classical radiomics methodology.
• Development of feature synthesis methods to work both on the classical methodology (genetic programming models, highly interpretable by the user) and on the methodology based on DL.
• Development of new models to improve the interpretability of the DL-based methodology.
• Development of descriptive analysis methods to work on the features generated by radiomics from medical images.

FUNDS

The development of this thesis was supported by:

• Regional Government of Andalusia, competitive grant to finance the pre-doctoral hiring of research personnel in training by agents of the Andalusian Knowledge System, file PREDOC_00369.
• Spanish Ministry of Science and Innovation and the European Regional Development Fund, under project PID2020-115832GB-I00.

PUBLICATIONS ASSOCIATED WITH THIS THESIS

INTERNATIONAL CONFERENCES

1. E. Almeda, J.M. Luna and S. Ventura. Radiomics Software Tools: A comparative Analysis on Breast Cancer. IEEE 36th International Symposium on Computer Based Medical Systems (CBMS) 2023, pp. XX-YY. 2023.