Liver Tumor Segmentation Using Classical Algorithms & Deep Learning

Abstract

Liver cancer is a common condition that traditionally required open surgery, posing a high risk of complications. Laparoscopic surgery has become increasingly popular, but comes with navigation challenges. The MedTech start-up Navari Surgical has developed a visualization solution using augmented reality, and this project aims to suggest a tumor segmentation method to support this solution. Previous studies have inspired this work to explore tumor segmentation utilizing different approaches, such as thresholding algorithms, active contour models, and a deep learning model utilizing the U-Net architecture. Thresholding methods uses pixel intensities, active contour models focuses on minimizing image energy, and U-Net models learn image features through training. For the U-Net models, variations in the learning rate, augmented data quantity, and loss functions are explored. The study utilizes the open-source LiTS dataset. The methods employ either liversegmented or cropped tumor images as inputs. Evaluation metrics include dice’s similarity coefficient (DSC) and recall, with a dataset of 107 images for evaluation of the classical algorithms, and 696 test images for the U-Net models. The obtained results demonstrate that thresholding algorithms with cropped input yield the highest DSC and recall values for the classical algorithms. The best performance was observed with cropped Multi Otsu (DSC: 0.435, recall: 0.605). For the U-Net models, increased augmented data, reduced learning rate, and more epochs resulted in improved performance. The best U-Net model achieved a DSC of 0.766 and a recall of 0.796. The discussion highlights challenges with algorithms designed for single tumor detection when evaluating datasets containing multiple tumors per image. Classical algorithms show a need for individualization for each scan, impacting automation and efficiency. Overfitting is a concern for the U-Net models, suggesting room for improvement. Further enhancements include pre and post-processing techniques, parameter variation, exploration of modified architectures, and utilization of 3D input data. In conclusion, U-Net demonstrated the best performance among the methods explored. However, its performance is not yet suitable for practical use, requiring further improvements. The recommendation for Navari is to continue to explore deep learning and U-Net for future advancements in tumor segmentation.