Browsing by Author "Honkamaa, Joel"

Now showing 1 - 5 of 5

ABC ja hierarkkiset mallit
(2016-08-30) Honkamaa, Joel
Sähkötekniikan korkeakoulu | Bachelor's thesis
Aligning images: Deformable registration of CT and Pseudo-CT images using unsupervised Deep Learning based algorithms
(2021-03-15) Honkamaa, Joel
Perustieteiden korkeakoulu | Master's thesis
Medical image registration is the process of algorithmically aligning medical images anatomically. Deformable registration seeks to find a nonlinear mapping between anatomic locations of different images. In the past few years Deep Learning has been successfully applied to the problem. In this work unsupervised Deep Learning based deformable registration is considered. The developed methodologies are applied to registration of computer tomography (CT) and artificial CT images created from magnetic resonance images. Regularization of the predicted mappings is an important aspect of deformable registration and diffeomorphisms, differentiable bijections with differentiable inverse, are often sought after. Very recently diffeomorphic registration frameworks have been applied to unsupervised Deep Learning registration and they have been shown to produce good results with very small run time. A common limitation in Deep Learning is the ability of a model to fit into a GPU memory. As a result patchwise approaches, where only sub-volumes of the whole data set are fed to the neural network at once are often employed. The approach introduces several problems unique for registration, especially ones related to the regularization of transformations. In this work an original framework for unsupervised Deep Learning registration based on image patches is introduced. The framework is shown to produce diffeomorphic transformations. Accuracy of the registrations is evaluated against a baseline and the method is shown to produce comparative results.
Deformation equivariant cross-modality image synthesis with paired non-aligned training data
(2023-12) Honkamaa, Joel; Khan, Umair; Koivukoski, Sonja; Valkonen, Mira; Latonen, Leena; Ruusuvuori, Pekka; Marttinen, Pekka
A1 Alkuperäisartikkeli tieteellisessä aikakauslehdessä
Cross-modality image synthesis is an active research topic with multiple medical clinically relevant applications. Recently, methods allowing training with paired but misaligned data have started to emerge. However, no robust and well-performing methods applicable to a wide range of real world data sets exist. In this work, we propose a generic solution to the problem of cross-modality image synthesis with paired but non-aligned data by introducing new deformation equivariance encouraging loss functions. The method consists of joint training of an image synthesis network together with separate registration networks and allows adversarial training conditioned on the input even with misaligned data. The work lowers the bar for new clinical applications by allowing effortless training of cross-modality image synthesis networks for more difficult data sets.
SITReg: Multi-resolution architecture for symmetric, inverse consistent, and topology preserving image registration
(2024-11) Honkamaa, Joel; Marttinen, Pekka
A1 Alkuperäisartikkeli tieteellisessä aikakauslehdessä
Deep learning has emerged as a strong alternative for classical iterative methods for de- formable medical image registration, where the goal is to find a mapping between the coordinate systems of two images. Popular classical image registration methods enforce the useful inductive biases of symmetricity, inverse consistency, and topology preservation by construction. However, while many deep learning registration methods encourage these properties via loss functions, no earlier methods enforce all of them by construction. Here, we propose a novel registration architecture based on extracting multi-resolution feature representations which is by construction symmetric, inverse consistent, and topology pre- serving. We also develop an implicit layer for memory efficient inversion of the deformation fields. Our method achieves state-of-the-art registration accuracy on three datasets. The code is available at https://github.com/honkamj/SITReg
Uncertainty estimation in medical image registration
(2023-08-21) Serkova, Elena
Perustieteiden korkeakoulu | Master's thesis
This thesis explores uncertainty estimation in deep-learning medical image registration. Image registration is an important part of medical imaging that aligns the images into one coordinate space to retrieve more information. Traditionally, image registration is performed manually. However, this approach is time-consuming and not accurate. Though software solutions improve registration, this process is still far from being accurate and effective. Deep learning models have shown promising results in automating this task, but lack of uncertainty estimation poses a limitation to implementing these models into clinical practice. Uncertainty indicates the level of confidence of the model in the output. This the- sis analyses the methods of uncertainty estimation and their suitability for medical image registration. The ensemble method is chosen for the project because this method can be applied to existing models, provides clear and transparent uncertainty estimation, and is easily scalable. Ensemble methods involve training several models with different initial states or architectures and combining their predictions to make more robust and accurate output. The ensemble of models allows uncertainty quantification through the measurement of standard deviation. As a result, it is possible to mark the regions with high uncertainty on a combined output deformed image. The results of uncertainty estimation are evaluated via several tests. Firstly, they are compared to the results of a probabilistic model. Though an ensemble and a probabilistic model return similar predictions, the ensemble is much better calibrated. In addition, a user study is conducted. In 70% of the cases, the opinion of the human experts coincides with the ensemble prediction. In conclusion, the results of tests conducted on the prediction of the ensemble show that ensemble methods are useful for uncertainty estimation in medical image registration. These methods can potentially enhance the accuracy and robustness of medical image registration.