University of Twente Student Theses
An assessment of the effect of multi-task learning in the recurrent inference machine for sparse magnetic resonance imaging reconstruction
Paquaij, MSc T.M. (2024) An assessment of the effect of multi-task learning in the recurrent inference machine for sparse magnetic resonance imaging reconstruction.
|
PDF
49MB |
| Abstract: | In recent decades, the increasing reliance on medical imaging for diagnostics has negatively impacted upon healthcare costs and efficiency. One of the reasons for this is that acquisition, reconstruction, image analysis, and diagnosis are traditionally separate steps, lengthening the diagnostic process. Accelerating these steps can save time and reduce costs. Accelerated magnetic resonance imaging (MRI) speeds up acquisition by minimising k-space samples, leading to aliased scans when linearly reconstructed. Deep learning methods, like recurrent inference machines (RIM), can reconstruct these scans faithfully using learned priors but currently exclude post-reconstruction tasks, missing the opportunity for improved performance through joint optimisation. This research examines the impact of multi-task approaches on the reconstruction and segmentation of sparse MRI data. The multi-task learning for accelerated-MRI reconstruction and segmentation (MTLRS) model is the baseline for this evaluation. This cascade-structured model incorporates multiple RIMs and Attention U-net modules informing each other through hidden states. Various multi-task approaches are assessed using uncertainty and predictive performance metrics, compared to MTLRS without a multi-task connection (JOINT). This study uses the Stanford Knee MRI with Multi-task Evaluation (SKM-TEA) dataset, which includes 155 3D Cartesian sampled Double Echo Steady State (DESS) knee scans with meniscus and cartilage segmentation labels, undersampled with an 8 × 2D undersampling Gaussian mask. Each multi-task approach is evaluated using reconstruction and segmentation metrics and an estimated quantitative T2 error metric. Uncertainty quantification is applied to identify variability within the unrolled network. The multi-task approach using a module with spatially adaptive semantic guidance (SASG) significantly improves reconstruction metrics compared to the JOINT approach. The Tukey honest significance difference (Tukey HSD) test demonstrated that, at a 95% confidence level, p < 0.001 for all metrics. Additionally, uncertainty estimation of intermediate predictions shows faster convergence to lower uncertainty with the SASG approach, indicating a positive impact on reconstruction. This research demonstrates that the SASG approach enhances the performance of the MTLRS model, which shows its validity across various unrolled multi-task reconstruction and segmentation networks. Furthermore, the proposed validation method could guide performance assessments of other unrolled network architectures. 2 |
| Item Type: | Essay (Master) |
| Faculty: | TNW: Science and Technology |
| Subject: | 31 mathematics, 54 computer science |
| Programme: | Biomedical Engineering MSc (66226) |
| Link to this item: | https://purl.utwente.nl/essays/103186 |
| Export this item as: | BibTeX EndNote HTML Citation Reference Manager |
Show download statistics for this publication
Show download statistics for this publicationRepository Staff Only: item control page