polito.it
Politecnico di Torino (logo)

Deep Learning Techniques for Breast Cancer Characterization in Magnetic Resonance Images

Angelo Laudani

Deep Learning Techniques for Breast Cancer Characterization in Magnetic Resonance Images.

Rel. Fabrizio Lamberti, Lia Morra. Politecnico di Torino, Corso di laurea magistrale in Ingegneria Informatica (Computer Engineering), 2021

[img]
Preview
PDF (Tesi_di_laurea) - Tesi
Licenza: Creative Commons Attribution Non-commercial No Derivatives.

Download (7MB) | Preview
Abstract:

Background: The aim of this thesis is to explore the solutions that Deep Learning techniques can offer in the field of Medical Imaging, in particular for breast cancer characterisation in magnetic resonance images. The thesis proposes the development of a Deep Learning architecture for a concrete problem such as the evaluation of pathological complete response (pCR) to neoadjuvant chemotherapy in breast cancer. Methods: The mpMRI dataset analysed includes 37 patients, each of whom underwent two studies: before and after 2 cycles of NAC. An index slice was extracted from each available sequence by an experienced radiologist. Pathological results were used as ground truth. The proposed architecture seeks to make the most of the multi-parametric nature of the dataset, extracting features separately from each of the available image modalities (DCE, DWI and T2). The resulting sub-sequences are used as input for a multi-task ensemble learning model that takes into account the different information carried by each of them, as well as the time dimension due to the two studies per patient. The use of a specific branch for each sub-sequence combined with the use of Grad-CAM aims to provide an additional level of interpretability to a model that starts directly from full slices. Results: Using 4-fold cross-validation, with each training set consisting of 28 patients and each validation set of 9, the mean area under the receiver operating characteristic (ROC) curve (AUC) of the model was 0.91, with a positive predictive value of 91.8% using all the available sub-sequences. Experiments with different configurations have shown that the combined use of all sub-sequences and both studies (pre-NAC and post-NAC) available per patient results in a model capable of better performance and generalisation. Conclusion: The work conducted in this thesis demonstrates the great potential of Deep Learning applied in this specific medical field, proposing a solution that achieves significant results in the use of an mpMRI dataset for early prediction of pCR to NAC, an area that is still little explored in the available literature and that could provide valuable information in a crucial task such as treatment prediction.

Relatori: Fabrizio Lamberti, Lia Morra
Anno accademico: 2020/21
Tipo di pubblicazione: Elettronica
Numero di pagine: 81
Soggetti:
Corso di laurea: Corso di laurea magistrale in Ingegneria Informatica (Computer Engineering)
Classe di laurea: Nuovo ordinamento > Laurea magistrale > LM-32 - INGEGNERIA INFORMATICA
Aziende collaboratrici: NON SPECIFICATO
URI: http://webthesis.biblio.polito.it/id/eprint/18163
Modifica (riservato agli operatori) Modifica (riservato agli operatori)