Medical image analysis and computing in breast cancer evaluation using mammography data

With more than two million new cases yearly, breast cancer is the most diagnosed cancer globally. Breast cancer diagnostics begin with patient symptoms, incidental findings, or results from screening mammography. Additional imaging can be conducted for a preliminary assessment.
In his doctoral dissertation, Antti Isosalo studied the application of computer vision and medical image analysis in breast cancer screening examination classification and the learning of different breast tissue types from direct digital mammograms and breast tomosynthesis data.
The first sub-study examined the maturity of machine learning methods for breast cancer screening examination classification. This study utilised a large, purpose-built dataset of nearly 50,000 screening studies collected from the Oulu University Hospital catchment area.
In the second sub-study of the thesis, the advantages of transfer learning in developing methods for cross-population mammography mass segmentation were evaluated. Two datasets with differing demographics were utilised in this sub-study.
The third sub-study investigated the applicability of deep learning for imaging phenotype classification from breast tomosynthesis data. This was addressed by collecting a suitable dataset and implementing a method to classify abnormality candidates from localised samples.
In the fourth sub-study, the suitability of local edge computing for augmenting the medical imaging workflow was investigated. This was done by defining technical requirements for a local edge computing platform and utilising a breast cancer detection method from the literature as a practical application.
This doctoral dissertation strengthened the perception that a strong pre-trained model is important when experimenting with a single-centre dataset, when the availability of malignant training examples is understandably limited. It was also stated in the thesis that developing reliable methods for automated breast cancer evaluation requires a large dataset. Moreover, the dataset should broadly cover various cancer subtypes, a requirement that can present a challenge in practice.
This doctoral dissertation was conducted at the Research Unit of Health Sciences and Technology at the University of Oulu in close collaboration with specialists from the Oulu University Hospital. The work was supported by the Jenny and Antti Wihuri Foundation, South Ostrobothnia Regional Fund of the Finnish Cultural Foundation, Thelma Mäkikyrö Foundation, Jane and Aatos Erkko Foundation, and Technology Industries of Finland Centennial Foundation.