Machine Learning and 3D Reconstruction Methods for Computational Pathology

Kimmo Kartasalo

Research output: Book/ReportDoctoral thesisCollection of Articles


Assessment of the microscopic anatomy of tissue samples forms the cornerstone of histopathological diagnostics. The current clinical practice is associated with challenges such as inter-observer variability and a global shortage of pathologists. Many fundamental aspects of pathology as a medical discipline have remained largely un- changed for decades, but the field is currently undergoing a transition into a digital discipline by replacing microscopes with whole slide scanners. Among other benefits, digital pathology unlocks the possibility of applying computational methods on the resulting image data. Some of the promises of computational pathology, such as improved efficiency and patient safety, take the advantages of digitization a step further, while others represent new types of analyses. This thesis focuses on two techniques in computational pathology: machine learning and 3D reconstructions.

Machine learning is a branch of computer science falling under artificial intelligence, which aims at emulating intelligent decision making. The field has progressed rapidly during the last decade due to the availability of larger datasets and improved computational resources. Deep learning in particular, representing a re- naissance of artificial neural network algorithms, has demonstrated unprecedented performance across a range of problems and is seen as revolutionary for histopathology. By streamlining the work of pathologists, machine learning tools could potentially mitigate the issues with the unsustainable workload and inter-observer variability, and even enable the discovery of new image-based prognostic markers.

Digital imaging also enables 3D histology, where serially sectioned tissue samples are reconstructed computationally. Conventionally, 2D tissue sections representing only limited cross-sectional views of the original 3D samples are used. Studying tis- sue in 3D holds potential for obtaining a more comprehensive view of normal and pathological processes where the spatial arrangement of different tissue structures or cell types is of relevance. Compared to direct 3D imaging using specialized instruments, computational reconstruction allows applying various histological and biochemical assays, while achieving subcellular resolution even for large tissue samples. The core methodological problem is how to align a sequence of 2D images to reconstruct a 3D volume without introducing distortions. Many algorithms have been proposed for the task, but an objective comparison of their performance has been lacking, complicating the application of 3D histology.

This thesis presents machine learning based systems for diagnostics of breast and prostate cancer, which represent a considerable fraction of all samples assessed in pathology departments worldwide. The system for assessing lymph node samples of breast cancer patients was based on extracting numerical features describing the tissue as input for random forest classifiers, and it was demonstrated to be capable of distinguishing between normal and metastatic tissue. This allows visually highlighting potentially malignant regions. The system for assessing prostate biopsies was based on deep neural networks and gradient boosted trees. It achieved clinically useful sensitivity and specificity in cancer detection, and cancer length estimates closely corresponding to those performed by a pathologist. In cancer grading, the system was comparable to a panel of specialized pathologists. This marks the first time that diagnostic performance comparable to specialists has been demonstrated on a large, clinically representative dataset of prostate biopsies.

The other two studies of the thesis present a framework for evaluating the quality of 3D reconstructions. The developed framework was applied to compare several publicly available algorithms and two commercial options. Moreover, the feasibility of automated hyperparameter tuning of reconstruction algorithms using Bayesian optimization was demonstrated for the first time. Algorithms relying on elastic transformation models capable of compensating for local tissue deformations were observed to achieve the most accurate reconstructions. Moreover, all of the studies in this thesis aimed at developing efficient ways of processing whole slide image data, resulting in a streamlined computational workflow utilizing parallel computing on graphics processing units on high-performance computer clusters.

Taken together, this thesis demonstrates that computational pathology techniques can achieve expert-level diagnostic performance, paving the way for the clinical adoption of such tools. The comparative results concerning 3D reconstruction algorithms highlight useful algorithmic features and hopefully promote further development of 3D histology from a prototype technique to a mainstream approach in biomedical research.
Original languageEnglish
Place of PublicationTampere
PublisherTampere University
ISBN (Electronic)978-952-03-1953-3
ISBN (Print)978-952-03-1952-6
Publication statusPublished - 2021
Publication typeG5 Doctoral dissertation (articles)

Publication series

NameTampere University Dissertations - Tampereen yliopiston väitöskirjat
ISSN (Print)2489-9860
ISSN (Electronic)2490-0028


Dive into the research topics of 'Machine Learning and 3D Reconstruction Methods for Computational Pathology'. Together they form a unique fingerprint.

Cite this