Abstract
Diminished localized deposition of VAT and SAT, homogeneity in the VAT and SAT density, augmented SAT volume, and a dispersed and heterogeneous distribution of abdominal muscle density are identified as risk factors for infectious pancreatic necrosis (IPN).
Introduction
Acute Necrotizing Pancreatitis (ANP) is a severe clinical condition characterized by pancreatic tissue death. Approximately 30% of ANP cases progress to Infectious Pancreatic Necrosis (IPN), a complication associated with mortality rates as high as 30%. Early identification of IPN is critical for clinical intervention. However, clinicians frequently rely on non-contrast CT scans, which often lack the resolution to distinguish early-stage necrosis compared to contrast-enhanced imaging.
Recent research suggests that body composition—specifically the distribution and quality of fat and muscle—plays a significant role in disease outcomes. This study utilizes artificial intelligence to extract "radiomic" features from standard CT scans to identify subtle patterns in body composition that predict the onset of IPN.
Methodology
The study employed a retrospective analysis of 315 patients diagnosed with ANP at Shengjing Hospital of China Medical University. The cohort included 101 patients with IPN and 214 with Non-Infected Pancreatic Necrosis (NPN).
The researchers developed a two-stage computational pipeline:
- Automated Segmentation: Utilizing the nnU-Net (no-new-Net) deep learning framework, the team automatically delineated four anatomical regions at the L3–L5 vertebral levels: Subcutaneous Adipose Tissue (SAT), Visceral Adipose Tissue (VAT), sacrospinalis muscle, and general abdominal musculature.
- Radiomics Extraction: Using the PyRadiomics package, 32 features were extracted per region, including first-order statistics (density distribution) and morphological characteristics (shape and geometry).
Key Insights: The Radiomic Signature of Infection
The analysis revealed distinct "hidden" patterns in the body composition of IPN patients that are not easily quantifiable by the human eye:
- Adipose Tissue Homogeneity: Contrary to expectations, IPN patients exhibited more uniform fat density and less localized accumulation (lower skewness) in both VAT and SAT. This suggests that a diffuse inflammatory process within the adipose tissue may precede systemic infection.
- Muscle Heterogeneity: While fat appeared more uniform, abdominal muscles in IPN patients were more "scattered" and heterogeneous, showing higher dispersion (Interquartile Range and Mean Absolute Deviation).
- Morphological Changes: Muscles in the IPN group were significantly more elongated and less spherical. These structural changes may reflect an adaptive response to pathology or intramuscular inflammation affecting the extracellular matrix.
- SAT Volume: Increased SAT volume, measured via 3D diameter, was identified as a risk factor, potentially linked to heightened systemic inflammatory responses.
Results and Clinical Outcomes
The nnU-Net model demonstrated high performance, achieving a 0.91 accuracy in segmenting the target vertebral regions. Statistical analysis confirmed several key predictors:
- VAT: Skewness ($p = 0.004$) and uniformity ($p = 0.036$) were highly significant.
- SAT: Skewness ($p = 0.023$) and maximum 3D diameter ($p = 0.044$) correlated with infection.
- Muscle: Interquartile Range ($p = 0.023$) and sphericity ($p = 0.010$) were primary indicators.
Clinically, the IPN cohort faced more severe outcomes, including higher rates of mechanical ventilation and significantly longer hospital stays (average 35.83 days vs. 19.71 days for NPN patients, $p < 0.001$).
Conclusion
This study demonstrates that AI-driven analysis of body composition on standard non-contrast CT scans can provide valuable predictive data for IPN. By identifying specific radiomic features in fat and muscle, clinicians can better assess risk in ANP patients. Future work will aim to integrate these metrics with direct pancreatic lesion analysis and multi-center data to enhance the generalizability and predictive power of the model.
Summary updated