AI-derived body composition parameters as prognostic factors in patients with HCC undergoing TACE in a multicenter study

Lukáš Müller, Aline Mähringer‐Kunz, Timo Alexander Auer, Uli Fehrenbach, Bernhard Gebauer, Johannes Haubold, Benedikt M. Schaarschmidt, Moon Kim, René Hosch, Felix Nensa, Jens Kleesiek, Thierno D. Diallo, Michel Eisenblätter, Hanna Kuzior, Natascha Roehlen, Dominik Bettinger, Verena Steinle, Philipp Mayer, David Zopfs, Daniel Pinto dos Santos, Roman KloecknerJHEP 2024 · DOI

TL;DR

This study leverages AI to automatically analyze body composition from CT scans of liver cancer patients undergoing TACE. They found that skeletal muscle volume, in particular, is a strong independent predictor of survival, suggesting its potential as an "opportunistic biomarker" for better patient stratification and treatment decisions.

What Problem Does This Paper Address?

Hepatocellular carcinoma (HCC), a common liver cancer, is often treated with transarterial chemoembolization (TACE). However, patients undergoing TACE are a very diverse group, making it challenging to accurately predict their prognosis and tailor treatment. Traditional risk assessment methods have limitations. The paper addresses this by exploring the utility of body composition assessment (BCA) parameters, derived from readily available CT scans, as prognostic indicators. Previous BCA studies in this area have often been monocentric, relied on manual measurements (which are time-consuming and prone to variability), and used 2D estimations rather than full 3D volumetry. This study aims to overcome these limitations by using a fully automated AI approach on a large, multicenter dataset to establish the prognostic value of BCA in HCC patients undergoing TACE.

Methods

The core of this study's methodology lies in its use of a fully automated, AI-based 3D volumetry pipeline for body composition assessment. This pipeline, built on a pre-trained convolutional neural network, analyzes pre-interventional CT scans to quantify various tissue compositions, including skeletal muscle volume (SM), total adipose tissue (TAT), visceral adipose tissue (VAT), subcutaneous adipose tissue (SAT), and intramuscular adipose tissue (IMAT).

Key aspects of their approach include:

  • Automated 3D Volumetry: Unlike previous studies relying on manual or semi-automatic measurements, this AI approach ensures high reproducibility and standardization.
  • Resampling to Standardized Slices: The algorithm automatically resamples CT slices to a uniform 5 mm thickness, addressing variations in acquisition parameters.
  • Federated Data Analysis: To maintain data privacy and comply with regulations, the AI model was deployed via Docker containers on local servers at each participating center. Data analysis occurred locally, with only aggregated results being shared.
  • Normalization: BCA parameters were normalized to the number of abdominal slices to account for variations in scan length and patient size.
  • Statistical Analysis: Both univariate and multivariate Cox proportional hazards regression models were used to assess the prognostic influence of BCA parameters and other clinical factors. Optimal cut-off values for BCA parameters were determined using the "survminer" package to create clinically meaningful risk groups.

Datasets

The study utilized a retrospective multicenter dataset comprising 754 treatment-naïve patients with HCC who underwent TACE at six tertiary care centers in Germany between 2010 and 2020. Pre-interventional CT scans were used for BCA analysis.

Key Insights and Ideas

  • Skeletal Muscle Volume as a "Super-Biomarker": The most striking finding is that skeletal muscle volume (SM) emerged as a consistent and independent prognostic factor for survival, even after accounting for other clinical variables. This highlights its significant role in patient outcomes.
  • AI for Scalability and Reproducibility: The successful implementation of a fully automated AI pipeline demonstrates its potential to make BCA analysis routine and accessible, overcoming the limitations of manual methods. This is crucial for widespread clinical adoption.
  • Opportunistic Biomarker Potential: BCA parameters, particularly SM, can be derived from CT scans already performed for treatment planning. This "opportunistic" nature makes them valuable without requiring additional patient procedures or costs.
  • Multicenter Validation is Key: By conducting the study across multiple centers, the researchers provide robust evidence that these AI-derived BCA findings are generalizable and not specific to a single institution's data or practices.
  • Beyond Tumor Burden: The study emphasizes that factors beyond tumor size and number, like body composition, are critical for predicting patient outcomes in HCC.

Key Results

  • Skeletal Muscle Volume (SM) is a Strong Predictor: In univariate analysis, low SM was associated with significantly impaired median overall survival (OS). Crucially, in multivariate analysis, SM remained an independent prognostic factor (p=0.039).
  • Other BCA Parameters Showed Mixed Results: While some other BCA parameters (e.g., TAT, SAT) showed univariate associations with survival, their predictive power diminished in multivariate analysis, underscoring SM's prominence.
  • Optimal Cut-offs Identified: The study identified optimal cut-off values for BCA parameters, allowing for the stratification of patients into high- and low-risk groups. Patients with low SM had a significantly worse median OS (15.8 months) compared to those with high SM (28.6 months, p <0.001).
  • BCLC Stage B Subgroup: Even within the BCLC stage B subgroup (where TACE is standard), SM remained an independent prognostic factor in multivariate analysis (HR 1.6, p=0.012).
  • Clinical Relevance: The findings suggest that integrating SM into existing scoring systems could enhance survival prediction and clinical decision-making for HCC patients undergoing TACE.

Limitations and Future Work

  • Retrospective Design: The study is retrospective, and the authors acknowledge the benefit of validating these findings in a prospective setting.
  • Heterogeneous Patient Population: While TACE is standard for BCLC stage B, the cohort included patients across BCLC stages 0, A, B, and C, reflecting real-world practice but potentially introducing variability.
  • Scanner Variability: The use of CT scans from different institutions and scanner types, despite normalization steps, could introduce some influence that is difficult to fully estimate.
  • Time-Dependent Risk Factors: The study focused on baseline BCA parameters. The authors note that BCA parameters can be time-dependent risk factors, and future research should investigate how oncological therapy influences these parameters over time.
  • Alternative Endpoints: The authors discuss the evolving landscape of HCC treatment with novel immunotherapies, suggesting that overall survival (OS) as the primary endpoint might be challenged. Future research could explore alternative endpoints.
  • Standardized Norm Values: A significant hurdle for clinical implementation is the lack of standardized norm values for BCA parameters, which the authors aim to address with their multicenter validation.

Summary updated

Abstract

Body composition assessment parameters, especially skeletal muscle volume, have been identified as relevant prognostic factors for many diseases and treatments. In this study, skeletal muscle volume has been identified as an independent prognostic factor for patients with hepatocellular carcinoma undergoing transarterial chemoembolization. Therefore, skeletal muscle volume as a metaparameter could play a role as an opportunistic biomarker in holistic patient assessment and be integrated into decision support systems. Workflow integration with artificial intelligence is essential for automated, quantitative body composition assessment, enabling broad availability in multidisciplinary case discussions.
BibTeX
@inproceedings{muller2024ai,
  author = {Lukáš Müller and Aline Mähringer‐Kunz and Timo Alexander Auer and Uli Fehrenbach and Bernhard Gebauer and Johannes Haubold and Benedikt M. Schaarschmidt and Moon Kim and René Hosch and Felix Nensa and Jens Kleesiek and Thierno D. Diallo and Michel Eisenblätter and Hanna Kuzior and Natascha Roehlen and Dominik Bettinger and Verena Steinle and Philipp Mayer and David Zopfs and Daniel Pinto dos Santos and Roman Kloeckner},
  title = {{AI-derived body composition parameters as prognostic factors in patients with HCC undergoing TACE in a multicenter study}},
  booktitle = {JHEP},
  year = {2024},
  doi = {10.1016/j.jhepr.2024.101125},
}
Figure