Artifact-resilient deep learning segmentation for enhanced structural and functional cardiac analysis

Cardiac imaging is essential in diagnosing heart disease. Imaging methods like x-rays, ultrasound, computed tomography (CT) and magnetic resonance imaging (MRI) can show the heart in great detail. Given AI’s superior learning capabilities, these imaging approaches can be greatly improved. It is critical to use methods that can withstand artifacts in cardiac MRI when segmenting the ventricle in order to guarantee accurate structural and functional investigation of those tissues. While many have focused on making algorithms better, very few have addressed the problem of artifacts in predictions and how they affect accuracy. To make earlier techniques more resistant to these phenomena, we investigate pretrained network fine-tuning in this work. Our suggested approach makes heavy use of data augmentations designed to resemble such objects. The findings significantly improved the baseline segmentations, with improvements up to 0.06 Dice score and 4 mm Hausdorff distance. Our proposed process consists of three steps: resampling and preprocessing of data and performing normalisation; loading and setting training parameters of a pretrained model; training the loaded model with the data, which is augmented by modifying it in various ways according to intensity and deformations.The 3D data is called volumes, and the pictures are the short-axis slices. We used 3D for data resampling, preprocessing, and normalization, and we used slices for data augmentation. During the segmentation job, our technique primarily focuses on fine-tuning an encoder. We resampled the volumes after orienting them to the canonical orientation. The area of interest was subsequently treated with a crop. After then, the data was divided into two parts: training and validation. Out of the total individuals, 20% were used for validation. Secondly, a histogram was generated from the training data and used to standardize the intensity. Using the aforementioned histogram, we standardize all sets. The data used in our research came from the Extreme Cardiac MRI Analysis Challenge, specifically the Respiratory Motion subset. The results showed a group of five people using the same four breathing intensity throughout the assessment. The ED and ES phases were recorded and labeled by a skilled radiologist for all individuals and breathing intensity.