The doctoral dissertation will be examined at the Faculty of Health Sciences at Kuopio Campus. The public examination will be streamed online.
What is the topic of your doctoral research? Why is it important to study the topic?
My dissertation represents a deep learning-driven approach for breast cancer risk and outcome prediction. This study aims to develop interpretable and evidence-based deep learning models to assist clinical practitioners in their regular workflow. Breast cancer is the most diagnosed cancer among women worldwide and remains a leading cause of cancer-related deaths. Despite improvements in screening and treatment, there are still critical gaps in early detection, accurate risk stratification, and personalized treatment planning. Clinicians often rely on multiple sources of information such as mammograms, pathology slides, and genetic testing, but integrating and interpreting these complex data modalities consistently is a real challenge.
My research addresses this challenge by building AI systems that make accurate predictions and provide interpretable insights that healthcare professionals can trust. By combining insights from imaging, histology, and genomics, this research aims to support more informed, data-driven decision-making in clinical practice. Ultimately, the goal is to move toward a more precise and personalised approach to breast cancer care that can lead to earlier detection, improved patient outcomes, and reduced burden on healthcare professionals.
What are the key findings or observations of your doctoral research?
This doctoral study provides valuable insights into how deep learning can be used to address critical challenges in breast cancer screening, diagnosis, and genetic risk analysis. By developing novel AI-driven methods tailored to specific clinical tasks, this research offers practical tools that enhance both accuracy and reliability in medical decision-making.
Key observations include the ability to estimate mammographic breast density, an established risk factor for breast cancer, with high precision using a multitask deep learning framework. The system demonstrated strong agreement with expert radiologists and supports additional diagnostic tasks such as BI-RADS classification and malignancy prediction. These findings highlight the potential of AI to improve population-level screening with consistent and scalable assessments.
Another major contribution lies in the development of interpretable deep learning models for digital pathology. The proposed solutions successfully automated the identification and quantification of diverse nuclei types in whole-slide images, supporting pathologists in analysing tumour composition more efficiently and objectively.
In addition, this study introduced a graph-based learning framework that captured complex interactions among breast cancer-associated genetic variants. The method provided interpretable outputs that align with biological understanding, offering a new avenue for refining genetic screening and risk stratification.
The key observations of this dissertation demonstrate that deep learning can play a transformative role in breast cancer care not by replacing human experts, but by enhancing their ability to make faster, more consistent, and evidence-based decisions. This work concludes that carefully designed, task-specific AI tools can bring tangible benefits to both clinical practice and biomedical research, paving the way toward more personalised, efficient, and trustworthy cancer diagnostics.
How can the results of your doctoral research be utilised in practice?
The results of this doctoral research can be utilised in several practical ways to enhance breast cancer care and clinical workflows.
The deep learning models developed for mammogram analysis, such as MTLSegNet and MV-DEFEAT, can be integrated into breast cancer screening programs to provide automated, consistent, and scalable assessments of breast density and malignancy risk. These tools can assist radiologists by reducing inter-observer variability and supporting more accurate risk stratification, particularly in high-volume screening settings.
In pathology, the models for nuclei segmentation and cell-type quantification, BayesNuSeg and CT-EMT, offer practical solutions for automating tissue analysis in digital pathology workflows. Their ability to accurately identify and count diverse cell types in whole slide images can support pathologists in routine diagnostics, tumour grading, and biomarker assessment while also reducing the time and effort involved in manual annotation.
The GenoGraph framework, developed for analysing genetic variants, provides a powerful tool for researchers and clinical geneticists. By identifying complex relationships among variants associated with breast cancer, it can support the discovery of novel genetic markers and improve the accuracy of genetic risk prediction models. This has potential implications for personalised screening, preventive strategies, and patient counselling.
The methods introduced in this research can be incorporated into clinical decision support systems, research pipelines, and educational platforms. Their emphasis on interpretability and reliability makes them suitable for real-world adoption, with the potential to improve diagnostic accuracy, reduce clinical workload, and contribute to more personalised and data-driven approaches to breast cancer care.
What are the key research methods and materials used in your doctoral research?
This research followed a problem-driven, application-oriented approach to develop and validate deep learning models for breast cancer risk assessment and outcome prediction across three key data modalities: mammographic imaging, digital pathology, and genetic variant analysis.
For mammographic imaging, a retrospective study was conducted using data from Kuopio University Hospital, where two expert radiologists helped generate 30,000 annotations from both institutional datasets and public repositories such as the Digital Database for Screening Mammography (DDSM). These mammograms, including mediolateral oblique (MLO) and craniocaudal (CC) views, were annotated with breast percentage density values, BI-RADS categories, and malignancy labels. In this doctoral research, two models are used: MTLSegNet, which utilises multitask learning for breast density estimation, and MV-DEFEAT, which employs evidential deep learning to classify mammogram findings while quantifying prediction uncertainty.
For digital pathology analysis, the study utilised hematoxylin and eosin (H&E) stained whole slide images (WSIs) of breast cancer tissue samples from open-access datasets, annotated for various nuclear cell types across tumour regions. Two models were developed: BayesNuSeg for Bayesian instance segmentation and CT-EMT for evidential multi-task learning, enabling automated nuclei segmentation and quantification to support reproducible cellular analysis while reducing pathologists' manual workload.
The genetic variant analysis component leveraged 1,500 genetic variants from institutional biobank repositories. The research implemented a novel graph-based contrastive learning framework called GenoGraph, which employed graph neural networks to model complex interactions between variants and uncover hidden genetic patterns associated with breast cancer susceptibility.
Throughout all studies, rigorous data preprocessing, model training, and hyperparameter tuning protocols were applied. Performance evaluation used standardised metrics including AUC, balanced accuracy, and uncertainty calibration scores. The research consistently prioritised interpretability, reproducibility, and clinical relevance to ensure the developed models could be effectively integrated into clinical workflows.
The doctoral dissertation of Naga Raju Gudhe, MSc, entitled A deep learning-driven approach for breast cancer risk and outcome prediction will be examined at the Faculty of Health Sciences. The Opponent in the public examination will be Associate Professor Pekka Ruusuvuori of the University of Turku, and the Custos will be Professor Arto Mannermaa of the 91. The public examination will be held in English.
