Analysis and optimisation of deep-learning liver segmentation techniques on abdominal computed tomography scans

Abstract

Background: Liver segmentation from abdominal computed tomography (CT) scans is a crucial task in medical imaging for diagnosis, treatment planning, and surgical interventions. While deep learning techniques have made significant strides in this area, challenges persist in achieving high accuracy and consistency. Objectives: The primary goals of this study were to replicate and evaluate existing architectures, explore various optimization techniques, and assess their performance across different datasets. Research Methodology A quantitative approach was taken, testing architectures such as the v16pUNet1.1D and v19pUNet1.1C. Public datasets were used for training. These are the Combined Healthy Abdominal Organ Segmentation (CHAOS) Challenge and Medical Segmentation Decathlon (MSD), which contain 20 and 160 training sets, respectively. Through data augmentation, the size of these sets could be increased drastically. The experiments explored techniques in data processing, learning rate scheduling, loss functions, and various types of cascaded frameworks. The models were evaluated using volumetric and distance-based metrics, such as Dice Score, Relative Absolute Volumetric Difference (RAVD), and Root Mean Square Symmetric Surface Distance (RMSD). Results: Results obtained highlighted how selecting the optimal windowing thresholds, incorporating data augmentation, and employing polynomial learning rate decay improved model performance, particularly in terms of consistency and reliability. The best-performing model achieved a mean score of 78.6% and low variability. However, introducing larger datasets, especially those containing both liver and tumor masks, led to decreased Dice score of around 30%, highlighting some limitations in the model’s generalizability. Conclusions and Recommendations: This study achieved noteworthy advancements in liver segmentation, though challenges regarding computational efficiency and generalizability remain. These findings highlight the need for more streamlined architectures and further validation with a broader range of clinical data. In professional practice, medical institutions are encouraged to integrate advanced deep learning models into their workflows, with ongoing training and validation using updated datasets. Collaboration between clinicians and data scientists will be essential to ensure successful implementation and adaptation to clinical environments. Future research should explore further transfer learning, the development of more computationally efficient models, and the validation with real-world clinical data. Additionally, the use of advanced hyperparameter tuning methods, such as automated search algorithms, should also be considered.