Master's Thesis

Characterization of Breast Tissue in Dynamic MR Scans

Final Thesis 19.44 MB Appendix 799.06 kB

Author of thesis: Bc. Filip Rudol

Acad. year: 2025/2026

Supervisor: Ing. Roman Jakubíček, Ph.D.

Reviewer: Ing. Madeleine Alabdah

Abstract:

Breast cancer is one of the most common oncological diseases among women. Breast magnetic resonance imaging enables the assessment of breast tissue and lesions, with dynamic contrast-enhanced MRI being the core breast protocol. The aim of this master's thesis is to develop a software tool for automated breast tissue characterization. The proposed pipeline includes the registration of dynamic scans and the segmentation of the breast, fibroglandular tissue, blood vessels, and lesions. This was achieved by implementing existing tools and models as well as training custom models. The nnU-Net framework is utilized in this work for the segmentation of all breast structures. Particular focus was placed on the segmentation of fibroglandular tissue and vessels, where the available vessel ground-truth annotations were refined, an inference ablation study was performed, and custom models were trained. The individual steps and the achieved results are evaluated and discussed.

Keywords:

breast tissue, breast cancer, dynamic contrast-enhanced MRI, fibroglandular tissue, vessels, image segmentation, nnU-Net

Date of defence

15.06.2026

Result of the defence

Defended (thesis was successfully defended)

znamkaAznamka

Grading

A

Process of defence

Student prezentoval výsledky své práce a komise byla seznámena s posudky. Doc. Šafránek položil otázku: Jaká je reprodukovatelnost softwaru? Jaké je state-of-the-art? V čem je vaše metoda přínosnější oproti publikovaným metodám? Prof. Provazník položila otázku: Co znamená pojem "složitější" případ, u kterého nebyly léze správně detekovány? Proč u tohoto subjektu model selhává? Student obhájil diplomovou práci a odpověděl na otázky členů komise a oponenta.

Language of thesis

English

Faculty

Department

Study programme

Biomedical Engineering and Bioinformatics (MPC-BTB)

Composition of Committee

doc. RNDr. David Šafránek, Ph.D. (předseda)
prof. Ing. Valentýna Provazník, Ph.D. (místopředseda)
Ing. Markéta Jakubíčková, Ph.D. (člen)
Ing. Roman Jakubíček, Ph.D. (člen)
Ing. Andrea Němcová, Ph.D. (člen)

Supervisor’s report
Ing. Roman Jakubíček, Ph.D.

Student pracoval velmi aktivně a samostatně, pravidelně konzultoval a prezentoval dosažené výsledky i problémy. V rámci řešení zvládl problematiku sám nastudovat a průběžně hledat řešení vzniklých situací. Samostatně si také vyhledal datasety, které dále rozšířil o vlastní manuální anotace, což je časově náročná činnost. Lze ocenit i zvládnutí práce s výpočetní infrastrukturou a celkovou přípravu a organizaci dat.
Text práce je přehledný, dobře strukturovaný a čitelný. Student vhodně popisuje použité datasety i jednotlivé kroky řešení. V praktické části implementoval a aplikoval řadu existujících modelů, které dále upravoval a trénoval. Jednotlivé kroky návrhu jsou logicky popsány a odůvodněny. Bylo provedeno velké množství experimentů včetně ablačních studií. Na jejich základě vznikl finální softwarový nástroj pro automatizovanou charakterizaci tkání, který je doplněn dokumentací a zveřejněn na GitHubu.
Práce přináší zajímavé výsledky s potenciálem dalšího rozpracování a možného publikačního využití. Menší výtky mám k prezentaci některých výsledků, kde by bylo možné zlepšit přehlednost či infografiku (např. zvýraznění hodnot v tabulkách, doplnění statistické významnosti nebo zobrazení rozptylu). V některých částech by bylo také možné uvést detailnější analýzu (např. statistická významnost využití segmentačních masek či vyhodnocení v rámci cross-validation).

Celkově se jedná o velmi kvalitní diplomovou práci. Zadání považuji za splněné a hodnotím stupněm A – 92 bodů. Points proposed by supervisor: 92

Grade proposed by supervisor: A

Reviewer’s report
Ing. Madeleine Alabdah


Chapters 1, 2, 4, and 5 are fully met: breast DCE-MRI and segmentation methods are surveyed, three public datasets are prepared and standardized, a complete pipeline is implemented, and the steps are tested and optimized through ablation studies. Chapter 6 (clinical applicability) is addressed only briefly. On Chapter 3, the student implicitly used subtraction maps — a basic form of parametric map — as input channels and visualization aids, and the ablation studies confirmed their value (the most important input for lesion segmentation). It would have strengthened this part to frame these as parametric maps and to briefly discuss the potential of further parametric maps as a possible extension.


The thesis is logically structured, and the theoretical part maps cleanly onto the practical work. The main text spans pages 14–54 (~41 pages). The content is dense and complete rather than padded, so the brevity reflects efficient writing; the deviation is nonetheless notable and partly reflects the undeveloped parametric-maps component. The introduction jumps from "DCE-MRI is important" to the objective without an explicit problem statement; In my opinion, adding the research gap would have strengthened the framing.


Formatting is professional and figures and tables are clear. Minor proofreading errors remain: "on on DCE-MRI" (p.14), "laciferous"/"lactiferous" (p.15), and "a integrated workflow" (p.52). Table 6.3 uses an inconsistent font size, and the dataset is referred to as both "ISPY2" and "I-SPY2." The thesis is written in fluent academic English, which is commendable.


Reference handling is a strength: citations are consistent and draw on current primary sources (Ronneberger, Isensee, Garrucho et al., Saha). The literature is current and relevant, and the Creative Commons figure is attributed correctly.


The registration module is validated subjectively and objectively (mutual information, paired Wilcoxon test); the FGT/vessel ablation studies are sound; data are split at the patient level to prevent leakage. The original contribution — iterative model-in-the-loop refinement of the vessel ground-truth annotations — is well-motivated and validated, with the model exceeding published baselines (FGT Dice 0.90 vs 0.86; vessels 0.73 vs 0.65). Failures are reported honestly, the code is public, and the pipeline is characterized for deployment (runtime, RAM).


The thesis presents a complete, reproducible, and technically well-executed pipeline for automated breast-tissue characterization, with a genuine original contribution in vessel ground-truth refinement and rigorous, honestly reported validation. It is weakened by the undeveloped parametric-maps component, and minor language and formatting lapses. The thesis meets the standard requirements for a Master's thesis.

I propose an overall grade of 92 points, corresponding to ECTS grade A. Topics for thesis defence:
  1. 1. Your subtraction images are themselves a basic parametric map, and your ablation studies show they carry significant information — yet the thesis does not frame them this way or explore richer parametric maps. Could you comment on why, and on whether quantitative kinetic maps over the full dynamic series might improve tissue characterization?
  2. 2. Your lesion ablation study suggests that combining the first post-contrast phase with subtraction images is promising, yet you retained the unmodified MAMA-MIA model. Why was fine-tuning not pursued as an intermediate option between using it as-is and training from scratch, and what would its computational cost have been?
Points proposed by reviewer: 92

Grade proposed by reviewer: A

Responsibility: Mgr. et Mgr. Hana Odstrčilová