Automated liver segmentation using Mask R-CNN on computed tomography scans

Abstract

Due to changes such as shape, border and density that occur in the slices of CT images, liver segmentation remains a difficult process. Compared to other segmentation methods, more successful segmentation results with deep learning models are general phenomenon. The Mask Regional-Convolutional Neural Networks (Mask R-CNN) framework is a method proposed for detecting key points on the image and segmentation. In this study, an automated computer-aided segmentation approach based on Mask R-CNN assisted by soft parameter selection for the region of interest (ROI) is proposed for high-accuracy segmentation and detection of the liver on CT images of the abdomen in three different datasets. Figure A. The methodology of the proposed Mask R-CNN model for the segmentation of the liver on CT images Purpose: Liver segmentation on slices of the scans acquired from abdomen region plays an important role in the clinical diagnosis and follow-up of the related diseases. Radiologists and physicians traditionally segment the liver or its region by manual segmentation. However, this process is highly time-consuming and the accuracy rate of the results may vary depending on the physician experience and skill. The aim of this study is to develop an automated computer-aided approach for high-accuracy segmentation and detection of the liver on CT images of the abdomen. Theory and Methods: In this study, a state-of-the-arts method based on Mask R-CNN is proposed that can assist physicians and specialists for segmentation of the liver on CT scans. It is observed that the proposed method is quite successful in the segmentation of the liver in experimental studies performed on a dataset of different sizes, with different scanning parameters and created specifically for this study, and two different publicly available datasets. In addition, the effectiveness and validity of the proposed method are verified by comparing the results of Mask R-CNN, supported by the proposed soft parameter selection for ROI, with the results of another popular segmentation algorithm, U-Net. Results: Experimental studies are conducted on three different liver CT image datasets, one of which is prepared specific for this study and two of them are public (Sliver07 and 3Dircadb), with both single and double GPU hardware structure. Thus, the change in segmentation performance depending on time is observed. The results obtained using the proposed method and the segmentation results realized by the specialist physician compared with parameters such as Dice similarity coefficient (DSC), Jaccard similarity coefficient (JSC), volumetric overlap error (VOE), average symmetric surface distance (ASD) and relative volume difference (RVD) metrics. In experimental studies carried out on liver CT dataset with the proposed Mask R CNN approach, DSC, JSC, VOE, ASD and RVD segmentation performance metrics are gained as 96.16%, 93.11%, 6.89%, 1.56 mm,-4.76%, respectively. Conclusion: With these results, it is seen that the proposed method in this study can be used as a secondary tool in the decision making processes of physicians for the segmentation of the liver.