Automatic Liver and Lesion Segmentation in CT Using Cascaded Fully Convolutional Neural Networks and 3D Conditional Random Fields (1610.02177v1)

Abstract: Automatic segmentation of the liver and its lesion is an important step towards deriving quantitative biomarkers for accurate clinical diagnosis and computer-aided decision support systems. This paper presents a method to automatically segment liver and lesions in CT abdomen images using cascaded fully convolutional neural networks (CFCNs) and dense 3D conditional random fields (CRFs). We train and cascade two FCNs for a combined segmentation of the liver and its lesions. In the first step, we train a FCN to segment the liver as ROI input for a second FCN. The second FCN solely segments lesions from the predicted liver ROIs of step 1. We refine the segmentations of the CFCN using a dense 3D CRF that accounts for both spatial coherence and appearance. CFCN models were trained in a 2-fold cross-validation on the abdominal CT dataset 3DIRCAD comprising 15 hepatic tumor volumes. Our results show that CFCN-based semantic liver and lesion segmentation achieves Dice scores over 94% for liver with computation times below 100s per volume. We experimentally demonstrate the robustness of the proposed method as a decision support system with a high accuracy and speed for usage in daily clinical routine.

• The paper presents a novel two-stage cascaded FCN architecture that first segments the liver and then targets lesions for enhanced accuracy.
• The integration of a 3D CRF refines segmentation outputs, addressing anisotropic resolution issues and achieving Dice scores above 94% for the liver and 82.3% for lesions.
• This robust method streamlines radiological workflows by providing consistent quantitative biomarkers for improved clinical diagnosis.

Automated Liver and Lesion Segmentation Using CFCNs and 3D CRFs

The paper presents a methodology for the automatic segmentation of liver and lesions in CT scans, an essential task for generating quantitative biomarkers for enhanced clinical diagnosis and computer-aided decision support. The research introduces a novel approach utilizing Cascaded Fully Convolutional Neural Networks (CFCNs) combined with 3D Conditional Random Fields (CRFs).

Approach and Methodology

The proposed method employs a two-stage cascaded FCN architecture. In the initial stage, a fully convolutional network segments the liver from the abdominal CT images to create a Region of Interest (ROI). Subsequently, a second FCN is applied solely to these liver ROIs to delineate lesions. This approach leverages the strength of CNNs in handling variations in image appearance to achieve resilient segmentation, even in low-contrast conditions typical in CT imaging. This is particularly valuable given the diversity in liver appearances and lesion contrasts.

To further refine the segmentation output and enhance spatial coherence, the authors incorporate a dense 3D CRF. By accounting for the spatial and intensity variations across the CT volume, this post-processing step addresses the anisotropic resolution issues inherent in 2D slice processing and boosts segmentation accuracy.

Experimental Results

The experimental evaluation was conducted on the 3DIRCADb dataset, which includes varied CT volumes with different aspects of liver contrast and lesion appearances. The evaluation metrics include the Dice coefficient, which showcases a remarkable improvement, achieving scores above 94% for liver segmentation. Moreover, the integration of the 3D CRF further enhances these results, reinforcing the robustness of the proposed pipeline.

Quantitative metrics indicate that the cascaded approach significantly enhances the segmentation quality over a single FCN. Furthermore, the cascade mechanism efficiently reduces false positive lesion detection, as demonstrated in the presented results where lesion segmentation Dice scores increased up to 82.3%.

Implications and Future Directions

The implications of this research are substantial, especially in terms of clinical applicability. By achieving high accuracy and speed, the method promises to streamline radiological workflows, potentially reducing the manual burden on radiologists while providing consistent and objective assessments. The adaptability of CFCNs to other medical imaging tasks also highlights its broader relevance and potential for deployment across various types of medical image analyses.

Looking ahead, there are multiple avenues for further research and development. Extending this methodology to include more sophisticated multiparametric imaging data could be an interesting trajectory, as well as adapting the model to real-time applications in clinical settings. Additionally, developing a generalized framework that can be fine-tuned for other organs and pathologies could prove valuable, facilitating new applications in the ever-evolving field of medical imaging and machine learning.

This work not only presents a robust and efficient method for liver and lesion segmentation but also contributes to the foundational infrastructure necessary for the ongoing development of automated medical image analysis solutions.