A flexible three-dimensional heterophase computed tomography hepatocellular carcinoma detection algorithm for
Chi-Tung Cheng1, Jinzheng Cai2, Wei Teng3
1Department of Trauma and Emergency SurgeryChang Gung Memorial Hospital at LinkouChang Gung UniversityLinkouTaiwan, Republic of China.
Hepatology Communications
|July 19, 2022
Summary
A novel deep learning algorithm, heterophase volumetric detection (HPVD), effectively detects hepatocellular carcinoma (HCC) across various computed tomography (CT) contrast protocols. This flexible tool shows promise for improving HCC surveillance and aiding clinical decisions.
Area of Science:
- Radiology and Medical Imaging
- Artificial Intelligence in Medicine
- Oncology
Background:
- Hepatocellular carcinoma (HCC) detection varies across different computed tomography (CT) contrast protocols, posing challenges for computer-aided detection (CADe) tools.
- Existing CADe tools often lack flexibility to adapt to diverse clinical imaging scenarios.
- There is a need for adaptable AI solutions to enhance HCC detection accuracy and consistency.
Purpose of the Study:
- To develop and evaluate a flexible deep learning algorithm capable of detecting HCC across multiple CT contrast protocols.
- To assess the performance of this algorithm compared to traditional CADe tools and human readers.
- To determine if a single, adaptable model can improve HCC detection in varied clinical settings.
Main Methods:
- Development of a flexible three-dimensional deep learning algorithm, heterophase volumetric detection (HPVD), designed to process various combinations of CT contrast phases (noncontrast, venous, dynamic contrast-enhanced).
- Training HPVD on 771 dynamic contrast-enhanced (DCE) CT scans for HCC detection.
- Evaluation of HPVD on a separate dataset including 164 positive cases and 206 controls, comparing its performance against six clinical readers.
Main Results:
- HPVD achieved area under the curve (AUC) values of 0.71 for noncontrast-only, 0.81 for noncontrast plus venous phase, and 0.89 for full DCE CT scans.
- At 80% sensitivity on DCE CT, HPVD demonstrated 97% specificity, comparable to physician performance.
- The flexible HPVD model outperformed less adaptable, nonheterophase detectors.
Conclusions:
- A single, flexible deep learning algorithm (HPVD) can be effectively applied to diverse HCC detection scenarios using varied CT contrast protocols.
- HPVD demonstrates potential as a valuable clinical aid for both at-risk and opportunistic HCC surveillance.
- The adaptability of HPVD suggests a path towards more generalized AI solutions in medical imaging for cancer detection.
