Deep Learning 2D image reconstruction algorithm in digital breast tomosynthesis

Digital Breast Tomosynthesis (DBT) is a new imaging modality that enhances breast cancer detection by having better resolution along the depth axis compared to digital mammography. However, it is challenging to obtain a correct 3D model of the breast from DBT imaging and standard reconstruction methods due to the limited number of views acquired. The reconstructions are also marred by dominant artifacts that make tissue identification difficult. To improve DBT reconstruction, Learned Experts Assessment-Based Reconstruction Network (LEARN) deep learning (DL) models are investigated in this work. Originally designed to enhance CT imaging with sparse data, we have redesigned and adjusted the LEARN algorithm to combat challenges of DBT reconstruction of 2D slices of digital breast phantoms. Performance of this technique was tested under two angular span conditions: 50° and 90°. The 90° case is not standard for routine DBT acquisition, but it was included as an experimental addition for estimation of the possible impact of additional projection data on reconstruction quality. One of the primary objectives of this study was to compare two DL networks: LEARN-CNN and LEARN-UNET. As LEARN networks, both integrate traditional iterative reconstruction methods with DL. However, their difference resides in the deep architecture: while LEARN-CNN applies only convolutional layers, LEARN-UNET employs an encoder-decoder network named as UNET. To better accommodate the varied geometric features of DBT compared to CT, the network was modified to incorporate a second channel so it could process data from DBT more effectively. Comparative visual assessment of the reconstructions employing the LEARN algorithm led to our main result: the LEARN algorithm clearly improves on traditional reconstruction. A quantitative assessment of fidelity and segmenting performance was considered via the Dice Coefficient, Precision, Recall, Structural Similarity Index and Relative Noise Magnitude Percentage, revealing that LEARN-UNET architecture consistently outperforms LEARN-CNN architecture across the two evaluated angular spans. Comparing the two angular span scenarios, the 90° case led to improvement in recalls, as expected due to the higher proportion of projection information; although this configuration does not reflect typical DBT acquisition constraints. Regarding clinical practicability, LEARN-UNET at 50° finds a better balance to attain high segmentation accuracy without sacrificing real-world DBT conditions. In order to further improve segmentation performance, preprocessing and post-processing techniques were explored, such as intensity thresholding and morphological filtering. The influence of skin tissue in segmentation was also examined, albeit with challenges related to over-segmentation, and an erosion-based post-processing technique was conducted to reduce false positives and improve segmentation stability. In summary, this thesis places LEARN-UNET at 50° as a promising reconstruction model of 3D DBT, which is capable of delivering high segmentation accuracy with adherence to realistic acquisition constraints. By integrating preprocessing, DL-based reconstruction, and post-processing optimizations, the study demonstrates that AI-based algorithms can significantly enhance the quality of DBT imaging and segmentation performance.