Estimation liver radiomics from postmortem CT: Development of interpretable models for postmortem interval estimation

Introduction Postmortem computed tomography (PMCT) is increasingly used in forensic investigations, offering a non-invasive and objective approach to estimating the postmortem interval (PMI). This study aimed to develop and externally validate radiomic models to distinguish deaths within versus beyond 24 h, using liver radiomic features from PMCT scans.. Methods A retrospective analysis was performed on 51 cadavers for model development and validated on 80 independent cases. In the training set, 173 PMCT scans across different PMIs were analyzed. The liver was manually segmented, and 40 radiomic features—statistical, morphological, and fractal—were extracted. Robustness to segmentation variability was assessed with autocontoured segmentations using the Intraclass Correlation Coefficient (ICC). PMI was dichotomized as ≤ 24 versus > 24 h. Univariate analyses identified predictive features, and logistic regression models were built from significant variables. Model performance was evaluated with receiver operating characteristic (ROC) curves, with sensitivity and specificity at the optimal threshold. Results Four features were significantly associated with PMI, with liver skewness emerging as the most predictive (p = 9.13 × 10−4) and robust (ICC = 0.75). A logistic regression model based on skewness achieved an AUC of 0.75 (95 % CI: 0.65–0.86) and 100 % specificity at the optimal threshold, reliably identifying deaths beyond 24 h. Adding a second feature did not improve performance (p = 0.54, DeLong test). External validation confirmed specificity of the skewness model (70 % at the optimal threshold). Conclusion Liver skewness extracted from PMCT shows potential as a biomarker for identifying deaths beyond 24 h, with performance confirmed on an independent cohort.