Developing a fully automated imaging biomarker for HCC risk assessment via MRI-based tumor segmentation and EPM.
Authors
Affiliations (11)
Affiliations (11)
- Department of Imaging Physics, The University of Texas MD Anderson Cancer Center, Houston, TX, USA. [email protected].
- Department of Imaging Physics, The University of Texas MD Anderson Cancer Center, Houston, TX, USA. [email protected].
- Department of Imaging Physics, The University of Texas MD Anderson Cancer Center, Houston, TX, USA.
- Department of Radiation Oncology Research, The University of Texas MD Anderson Cancer Center, Houston, TX, USA.
- Department of GI Radiation Oncology-Research, The University of Texas MD Anderson Cancer Center, Houston, TX, USA.
- Department of Molecular & Cellular Oncology, The University of Texas MD Anderson Cancer Center, Houston, TX, USA.
- Department of Gastroenterology, Houston Methodist Hospital, Houston, TX, USA.
- Department of Radiology, Houston Methodist Hospital, Houston, TX, USA.
- Department of Epidemiology, The University of Texas MD Anderson Cancer Center, Houston, TX, USA.
- Department of Hepatology, Baylor College of Medicine, Houston, TX, USA.
- Department of GI Radiation Oncology, The University of Texas MD Anderson Cancer Center, Houston, TX, USA.
Abstract
This study investigates the feasibility of using automated tumor segmentation as the region of interest for early detection of hepatocellular carcinoma (HCC) using magnetic resonance imaging (MRI). Enhancement Pattern Mapping (EPM) is used as a voxel-wise imaging biomarker within the region of interest once segmentations are generated. We implement PocketNet, a lightweight convolutional neural network, to segment liver tumors and evaluate performance. To contextualize model accuracy, we estimate the theoretical upper bound of segmentation performance through inter-observer comparisons of manual annotations. Imaging-derived features from automated segmentations are analyzed using XGBoost to assess their predictive value. Results show that automated segmentation approaches upper-bound performance for larger tumors but underperforms on smaller lesions. However, EPM features derived from automated masks demonstrate comparable predictive power to those from manual segmentations, indicating that the EPM features are stable with respect to segmentation inaccuracies. Results demonstrate that a fully automated imaging biomarker may be developed as a clinical utility for HCC risk assessment.