Advanced methods of medical image analysis for breast cancer risk prediction

Klaneček, Žan (author), Jeraj, Robert (mentor)

Breast cancer screening programs are generally effective; however, it remains unclear whether the uniform screening interval applied to all women is optimal. Personalized screening—tailoring the frequency to an individual's breast cancer risk (BCR)—may provide a better solution. For the realization of personalized BCR screening, this thesis focuses on developing a deep learning (DL)-based BCR prediction model specifically optimized for the Slovenian population, leveraging all relevant information available from routine 2D mammograms within screening programs. Initially, the state-of-the-art BCR prediction model, MIRAI, was validated on Slovenian screening data, achieving performance comparable to that reported at other screening centers. The evaluation also revealed that MIRAI occasionally bases its predictions on the pectoral muscle (PM) region in mammograms without clear clinical justification. To address this, we hypothesized that excluding the PM region during model training would improve predictive performance. Consequently, a PM segmentation model incorporating an uncertainty-based flagging mechanism using Monte Carlo dropout was developed and validated. Fine-tuning experiments confirmed our hypothesis: BCR prediction models trained on mammograms with the PM region removed demonstrated superior discrimination in predicting 1–5-year BCR on Slovenian data. Recognizing that these models could be sensitive to changes introduced by the physics-based principles of image acquisition and biological variations, we systematically quantified their impact through controlled experiments simulating realistic image alterations. Finally, to increase trust and clinical acceptance, a longitudinal interpretability study examined how the model’s reliance on the cancer-affected breast side evolves over time. In summary, the findings and technical innovations presented in this thesis contribute significantly to advancing personalized breast cancer screening in Slovenia. However, prior to the implementation of DL-based BCR screening in clinical practice, results from prospective clinical trials are needed.

• Language: English
• Keywords: breast cancer, breast cancer risk, artificial intelligence, machine learning, deep learning, uncertainty estimation, sensitivity, interpretability, segmentation, pectoral muscle, personalized screening, mammography
• Work type: Doctoral dissertation
• Typology: 2.08 - Doctoral Dissertation
• Organization: FMF - Faculty of Mathematics and Physics
• Year: 2025
• PID: 20.500.12556/RUL-170501
• COBISS.SI-ID: 244299523
• Publication date in RUL: 07.07.2025

Secondary language

• Language: Slovenian
• Title: Napredne metode analize medicinskih slik za napovedovanje ogroženosti raka na dojki

Abstract

Presejalni programi za raka dojk so na splošno učinkoviti, vendar še vedno ni jasno, ali je enoten interval presejanja, ki velja za vse ženske, optimalen. Rešitev bi lahko predstavljalo personalizirano presejanje, pri čemer bi se interval presejanja za vsako žensko prilagodil glede na njeno individualno ogroženost za raka dojk (ORD). V smeri potencialne uresničitve personaliziranega presejanja je bil primarni cilj te doktorske disertacije razvoj naprednega modela za napovedovanje ORD, ki temelji na metodah globokega učenja (GU) in je posebej optimiziran za slovensko populacijo, pri čemer izkorišča vse razpoložljive informacije iz rutinskih 2D mamogramov, pridobljenih v okviru presejalnega programa. Najprej smo na podatkih iz slovenskega presejalnega programa DORA validirali model MIRAI, ki velja za najboljši javno dostopen model za napovedovanje ORD, in potrdili učinkovitost, primerljivo z drugimi presejalnimi programi. Pri analizi smo ugotovili, da se MIRAI pri napovedovanju ORD osredotoča tudi na klinično nerazložljiva območja v pektoralni mišici (PM). Zato smo postavili hipotezo, da izključitev območja PM med učenjem modela na mamografskih slikah lahko privede do boljšega napovedovanja ORD. Za odstranjevanje PM iz mamogramov smo razvili segmentacijski model z integriranim mehanizmom označevanja slabih segmentacij, ki temelji na oceni negotovosti s pomočjo metode Monte Carlo. Rezultati učenja in validacije modela na mamogramih z odstranjeno PM so pokazali boljše rezultate za napovedovanje ORD s prediktivnim horizontom do 5 let. Ker so ti modeli lahko občutljivi na realistične spremembe, ki izhajajo iz fizikalnega ozadja zajemanja slik in bioloških variacij, smo njihov vpliv sistematično ovrednotili s pomočjo kontroliranega spreminjanja vhodnih slik. Za povečanje zaupanja in klinične sprejemljivosti smo izvedli tudi longitudinalno interpretabilnostno študijo, v kateri smo preučili, kako se sčasoma spreminja odvisnost modela glede na dojko, v kateri se bo sčasoma razvil rak. Ugotovitve in tehnične inovacije, predstavljene v tej disertaciji, pomembno prispevajo k razvoju personaliziranega presejanja raka dojk v Sloveniji. Pred dejansko implementacijo modelov za napovedovanje ORD, ki temeljijo na GU, pa so potrebni rezultati prospektivnih kliničnih študij.

• Keywords: rak dojk, ogroženost za raka dojk, umetna inteligenca, strojno učenje, globoko učenje, ocenjevanje negotovosti, senzitivnost, interpretabilnost, segmentacija, pektoralna mišica, personalizirano presejanje, mamografija