Abstract
A DEEP LEARNING-BASED HEPATOCELLULAR CARCINOMA STAGING SYSTEM BY MULTI-PHASE COMPUTED TOMOGRAPHY
Background:
Hepatocellular carcinoma (HCC) is a leading cause of cancer. The Barcelona Clinic Liver Cancer (BCLC) staging system is widely used to diagnose and manage HCC. However, currently, automated image-based staging algorithms are lacking. This study aims to automate the BCLC staging process using deep learning techniques on multi-phase abdominal computed tomography (CT) images to enhance efficiency with good accuracy.
Methods:
This study utilizes two open datasets: Multimodality annotated HCC cases with and without advanced imaging segmentation (HCC-TACE-Seg) dataset and The Medical Segmentation Decathlon (MSD) Liver dataset. A novel two-stage deep learning pipeline for automatic BCLC staging was generated. We employ an automatic segmentation model in the proposed pipeline to locate the liver and possible lesions. Then, a 3D image classification model is applied to the liver to identify the lesions for BCLC staging. We adopt the nnU-Net as our segmentation model backbone. The segmentation model is trained and evaluated on the combined HCC-TACE-Seg and MSD datasets. For the BCLC staging model, we adopt the 3D ResNet as our model backbone and train on the HCC-TACE-Seg dataset. Furthermore, we proposed three inter-phase fusion techniques (early fusion, intermediate fusion, and late fusion) to utilize the information from the multi-phase CT images.
Results:
Our liver segmentation results were consistent with existing algorithms in accurately delineating the liver region (dice score 0.95), while liver tumor segmentation ranked above average among evaluated methods (dice score 0.64 vs. 0.54). We compared the multi-phase models with single-phase imaging for BCLC staging. Among the proposed multi-phase models, late fusion performed the best, and outperformed single-phase ones with an overall accuracy of 77.78% (vs. 68.89% in the pre-contrast phase and PV phase, 73.33% in A phase), demonstrating the importance of inclusion of different phases of CT images for accurate staging. The best prediction accuracy for BCLC stages A, B, and C were 93.33%, 73.33%, and 66.67% respectively.
Conclusion:
Our numerical experiments demonstrate that the multi-phase staging models outperform the single-phase one and image processing-based method with an accuracy of 77.78%. Overall, the proposed pipeline can aid the radiologist for the BCLC staging process in the future.