Estimating deviation from Canonical Orientation for CT Head images

• Published in: 2024 IEEE International Symposium on Biomedical Imaging (ISBI) pp. 1 - 4
• Main Authors: Anand, Deepa, Singhal, Vanika, Mullick, Rakesh, Patil, Uday, Dutta, Sandeep, Das, Bipul
• Format: Conference Proceeding
• Language: English
• Published: IEEE 27-05-2024
• Subjects: Accuracy; Alignment correction; Annotations; Computed tomography; deep learning; Head; Prediction methods; Three-dimensional displays; Visualization
• ISSN: 1945-8452
• DOI: 10.1109/ISBI56570.2024.10635388

Aligning medical images with standard visualization planes is crucial for accurate diagnosis. The literature suggests two categories of methods for correcting misaligned medical images: registration techniques, which are less reliable when there is a significant rotation angle or the presence of pathology or foreign objects, and landmark-based methods, which require a substantial number of landmark annotations and their performance is contingent on the accuracy of the landmark segmentation. In this work, we evaluate different direct transformation methods, which predict the transformation parameters for volumetric image alignment. In the first method, we present a modified version of Iterative Transformation Networks (ITN) for estimation of Euler angles along with the directions. In the second method we propose for direct transformation parameter prediction, works on 2D projections of 3D volume, making it faster than traditional slice-based and 3D processing-based methods. These approaches significantly reduce the manual effort required to generate ground truth, as it only requires a small set of correctly oriented images to train the model. Additionally, we also explore the utility of a task-specific self-supervised pretraining, followed by fine-tuning on a smaller datasets for the 2D projection method. We evaluate the various direct prediction methods on a test set of head CT images and demonstrate the efficacy of direct prediction methods over traditional image alignment methods despite requiring minimal annotation efforts. An average error of ≤5° is observed with proposed methods for all the orientations.