MRI-based radiomics-deep learning model for preoperative pathogen prediction in perianal abscesses.
Authors
Affiliations (6)
Affiliations (6)
- Department of Radiology, Third Affiliated Hospital of Soochow University, Changzhou, Jiangsu, China.
- Department of Radiology, Pudong Gongli Hospital, Shanghai University of Medicine and Health Sciences, Shanghai, China.
- School of Gongli Hospital Medical Technology, University of Shanghai for Science and Technology, Shanghai, China.
- Department of Traditional Chinese Medicine Anorectal Surgery, Pudong Gongli Hospital, Shanghai University of Medicine and Health Sciences, Shanghai, China.
- Department of Clinical Laboratory Medicine, Pudong Gongli Hospital, Shanghai University of Medicine and Health Sciences, Shanghai, China.
- Shanghai Health Commission Key Lab of Artificial Intelligence (AI)-Based Management of Inflammation and Chronic Diseases, Sino-French Cooperative Central Lab, Pudong Gongli Hospital, Shanghai University of Medicine and Health Sciences, Shanghai, China.
Abstract
This study aimed to develop a hybrid model combining MRI-based radiomics, deep learning, and clinical variables for preoperative differentiation of <i>Escherichia coli</i> from non-<i>Escherichia coli</i> pathogens in perianal abscesses. A retrospective series of 215 patients with culture-confirmed perianal abscesses (119 <i>Escherichia coli</i>, 96 non-<i>Escherichia coli</i>) was analyzed. Preoperative MRI data were collected, with radiomic and deep learning features extracted from T1-weighted imaging (T1WI), T2-weighted imaging (T2WI), and fat-suppressed T2WI (FS-T2WI) sequences. Radiomic feature selection was performed using univariate <i>t</i>-tests, Pearson correlation, and least absolute shrinkage and selection operator (LASSO) regression. Clinical and MRI data were screened through univariate and multivariable logistic regression. The MRI signature was derived by averaging the probabilities from a logistic regression model (radiomics) and a k-nearest neighbors classifier (deep learning). A hybrid logistic regression model integrated the MRI signature with clinical predictors to create a nomogram. Model performance was evaluated using receiver operating characteristic (ROC) analysis, calibration curves, and decision curve analysis (DCA). <i>Escherichia coli</i> accounted for 57.63% of all strains. Gender (odds ratio [OR] = 0.456, <i>p</i> = 0.001) and diabetes (OR = 9.207, <i>p</i> < 0.001) were significant independent predictors. The nomogram achieved an area under the curve (AUC) of 0.885 in the testing set, outperforming the MRI signature alone (AUC = 0.860), with accuracy of 0.815, sensitivity of 0.818, and specificity of 0.812. Calibration curves showed good agreement between predicted and observed outcomes, while DCA demonstrated superior clinical utility. The hybrid model, utilizing preoperative multi-sequence MRI, noninvasively identifies <i>Escherichia coli</i>, the predominant pathogen in perianal abscesses, offering significant clinical potential to transition from empirical antibiotic regimens to microbiology-guided precision strategies.