High-Precision 3D Spleen Segmentation for Surgical Planning Using Deep Learning

Abstract

Medical imaging has become an indispensable tool in modern healthcare, facilitating accurate visualization of internal anatomical structures for diagnosis, treatment planning, and surgical procedures. Among these, segmentation of organs such as the spleen is particularly critical for pre- surgical workflows, including radiation therapy planning, volume estimation, and surgical navigation. The spleen, an organ of considerable clinical significance, often requires precise delineation for managing trauma or pathologies like lymphoma. However, spleen segmentation from computed tomography (CT) images is challenging due to anatomical variability, low tissue contrast, and the presence of imaging artifacts. Traditional segmentation methods, such as thresholding and active contour models, often fail to generalize across diverse datasets, necessitating advanced approaches. This study aims to address these challenges by developing a robust segmentation framework based on a 3D U-Net architecture. The primary objectives include achieving high segmentation performance on the Medical Segmentation Decathlon (MSD) Task 09 spleen dataset, evaluating the model's generalization capabilities, and demonstrating its clinical relevance for pre-surgical workflows. The methodology involves a systematic preprocessing pipeline to normalize and augment the CT volumes, followed by training a 3D U-Net model using Dice loss and the Adam optimizer over 600 epochs. The segmentation performance was evaluated using metrics such as Dice similarity coefficient and validation loss to ensure accuracy and reliability. The proposed framework achieved remarkable results, with a training Dice similarity coefficient of 0.9582, a validation Dice similarity coefficient of 0.9494 and a test Dice Similarity coefficient of 0.9483. These results highlight the model's strong generalization ability and effectiveness in addressing challenges such as noisy imaging and anatomical variability. Qualitative evaluations further confirm the precision of the segmentation, with predicted masks showing high alignment with ground truth annotations. In conclusion, this study presents a significant contribution to automated medical image segmentation by leveraging the power of deep learning. The results not only validate the efficacy of the proposed framework but also lay the groundwork for extending this approach to multi-organ segmentation tasks and real-time clinical workflows. Future research directions include integrating multi-modal imaging data, improving computational efficiency, and validating the framework in clinical settings to ensure broader applicability and impact.