A DWI histogram-based nomogram for preoperative prediction of pathogens and antibiotic resistance in perianal abscesses

Abstract

Purpose

To develop nomograms based on diffusion-weighted imaging (DWI) histogram parameters and clinical features to preoperatively predict pathogen type and extended-spectrum β-lactamase (ESBL) infection in perianal abscesses.

Methods

We retrospectively analyzed 157 surgically confirmed patients, stratified by pathogen type (Escherichia coli, n = 110; Klebsiella pneumoniae, n = 47) and ESBL test results (ESBL-negative, n = 91; ESBL-positive, n = 30). Ninety-seven apparent diffusion coefficient (ADC) histogram parameters were extracted. Histogram features selected using least absolute shrinkage and selection operator (LASSO) regression, together with clinical variables identified by univariate logistic regression, were incorporated into multivariate logistic regression models to construct nomograms. Internal validation used 1,000 bootstrap resamples. Model performance was assessed using the area under the receiver operating characteristic curve (AUC), the Hosmer–Lemeshow test, calibration curves, and decision curve analysis (DCA).

Results

The pathogen discrimination model, integrating an ADC-derived composite score (ADC_Score) based on 20 retained histogram features with age, diabetes, and hypertension, achieved an AUC of 0.897, sensitivity of 0.872, and specificity of 0.809. The ESBL prediction model, incorporating ADC_Score based on 13 retained features together with white blood cell count (WBC) and age, yielded an AUC of 0.823, sensitivity of 0.867, and specificity of 0.659. Calibration curves and the Hosmer–Lemeshow test indicated good agreement between predicted and observed probabilities, and DCA suggested potential net benefit for both models within the internally validated cohort.

Conclusion

DWI histogram-based nomograms demonstrated promising performance for pathogen prediction in perianal abscesses, while the incremental value for ESBL prediction was limited. These models represent an internally validated development study and require external validation before clinical application.

Supplementary Information

The online version contains supplementary material available at 10.1007/s00384-026-05109-7.

Introduction

Perianal abscess (one form of anorectal abscess) is a common infectious condition in colorectal surgery that typically originates from infection of the anal glands [1]. Patients usually present with perianal pain, swelling, and erythema, while purulent discharge may occur in severe cases [1]. The condition affects individuals of all ages but is most frequently observed in men aged 20 to 40 years [2]. A Swedish population-based cohort study reported an annual incidence rate of 16.1 per 100,000 population [3]. Although relatively uncommon, the disease causes considerable discomfort and significantly impairs quality of life and psychological well-being [4].

Currently, surgical incision and drainage with fistula exploration constitute the standard therapeutic approach for perianal abscesses [5]. In cases of extensive infection, systemic inflammatory response, or heightened risk of postoperative complications, adjunctive antibiotic therapy is often necessary [6]. Accurate identification of the causative pathogens and their resistance profiles is therefore critical for guiding perioperative antimicrobial management [6]. Previous studies have shown that Gram-negative bacilli, particularly Escherichia coli (E. coli) and Klebsiella pneumoniae (K. pneumoniae), are the predominant pathogens in perianal abscesses [7]. These organisms are also major producers of extended-spectrum β-lactamases (ESBLs) [8]. ESBL-producing strains often display multidrug resistance, creating substantial challenges for clinical management [8]. Accordingly, distinguishing infections caused by these predominant pathogens and assessing the risk of ESBL infection are essential for tailoring effective preoperative antimicrobial therapy [9].

In recent years, widespread antibiotic misuse has contributed to the global rise of ESBL-mediated resistance, posing major challenges to empirical therapy [10]. ESBL-producing organisms compromise the efficacy of many β-lactam agents; consequently, carbapenems often remain key therapeutic options for ESBL-producing infections, yet their overuse may accelerate the emergence of carbapenem-resistant strains [11, 12]. Early identification of patients at risk for ESBL-positive infections is therefore crucial for promoting rational antibiotic use and improved treatment planning [13, 14]. However, ESBL detection still relies largely on postoperative pus culture and susceptibility testing, which are time-consuming and may delay timely intervention [15]. Accordingly, effective preoperative prediction models may help support early risk stratification and inform empirical therapy.

Magnetic resonance imaging (MRI) is the preferred modality for evaluating perianal abscesses and fistulas, owing to its superior soft-tissue resolution and multiplanar imaging capabilities that enable comprehensive evaluation of anorectal anatomy and disease extent [16–18]. Diffusion-weighted imaging (DWI) and its derived apparent diffusion coefficient (ADC) provide quantitative insights into tissue water diffusivity, thereby reflecting underlying pathophysiological alterations [19]. Conventional ADC analysis is usually based on mean values, which may not sufficiently capture intralesional heterogeneity [20]. In contrast, histogram analysis examines the distribution and variation of voxel intensities within the region of interest (ROI), providing a more nuanced depiction of tissue complexity [20]. Recent studies have demonstrated the utility of ADC histogram analysis in tumors and infectious lesions, showing that its parameters can effectively differentiate lung cancer from pulmonary abscesses and mycobacterial infections [21–23].

However, effective preoperative tools for simultaneously predicting pathogen type and ESBL infection risk in perianal abscesses remain limited. By capturing microscopic tissue architecture and intralesional heterogeneity, DWI-based histogram analysis may offer a useful imaging approach for such predictions [20]. Therefore, this study aimed to develop and internally validate nomograms combining ADC histogram features with clinical variables for noninvasive preoperative prediction of pathogen type and ESBL infection risk, with the potential to support perioperative risk stratification and inform empirical antibiotic considerations pending further external validation.

Materials and methods

Patients

This retrospective study consecutively enrolled 215 patients with surgically confirmed perianal abscesses between August 2019 and March 2025. The inclusion criteria were as follows: (1) age ≥ 18 years; (2) a clinical diagnosis of perianal abscess; and (3) availability of preoperative pelvic multi-b-value DWI. The exclusion criteria were as follows: (1) identification of pathogens other than E. coli or K. pneumoniae; (2) poor image quality caused by MRI artifacts or other factors; and (3) incomplete or missing clinical data. According to the predefined exclusion criteria, patients with incomplete or missing clinical variables were excluded at the enrollment stage. All imaging-derived variables, including ADC histogram features, lesion volume, and derived ADC scores, were complete in the final study cohort; therefore, no imputation procedures were applied in the subsequent analyses. Based on these criteria, 157 patients were ultimately included in Task 1 (pathogen identification) and classified into the E. coli group (Group 1, n = 110) or the K. pneumoniae group (Group 2, n = 47). For Task 2 (ESBL prediction), 36 patients without ESBL testing were excluded, resulting in a final cohort of 121 patients, who were further divided into the ESBL-negative group (Group 3, n = 91) and the ESBL-positive group (Group 4, n = 30). ESBL testing was not available for these 36 patients because the corresponding results could not be retrieved from the electronic medical record system in this retrospective study. A detailed enrollment flowchart is presented in Fig. 1.

Fig. 1.

Flowchart of patient selection and study design. E. coli = Escherichia coli; K. pneumoniae = Klebsiella pneumoniae; ESBL = extended-spectrum β-lactamase

Pathogen identification and ESBL testing

After surgical incision and drainage, approximately 2–5 mL of pus was collected using a sterile syringe. Specimens were inoculated onto blood agar (BAP) and MacConkey agar (MAC) plates (Yihua Biotech, China) and incubated at 35 °C in an atmosphere containing 5–10% CO₂ for 18–24 h. Following incubation, bacterial colonies were Gram-stained and examined under light microscopy, including oil immersion, for preliminary morphological assessment. Definitive identification was conducted using matrix-assisted laser desorption/ionization time-of-flight mass spectrometry (MALDI-TOF MS; EXS1600 series, Zhongyuan Huiji, China), with spectral profiles compared against a proprietary database to determine the causative pathogens.

ESBL screening and antimicrobial susceptibility testing were carried out using the VITEK 2 Compact automated microbial identification system (bioMérieux, France). Susceptibility to third-generation cephalosporins (e.g., ceftriaxone, ceftazidime) was initially evaluated. Strains resistant to these agents but showing intermediate or susceptible phenotypes when combined with a β-lactamase inhibitor (e.g., clavulanic acid) were classified as ESBL producers (ESBL-positive).

MRI acquisition

All MRI examinations were performed on a 3.0 T scanner (Vantage Titan, Canon Medical Systems, Japan) with a 16-channel abdominal phased-array coil. Patients were imaged in the supine position without bowel preparation or administration of antispasmodics. The standardized protocol included multi-b-value DWI using a single-shot echo-planar imaging sequence. Seven b values were acquired: 0, 400, 600, 800, 1000, 1500, and 1600 s/mm². Imaging parameters were as follows: repetition time (TR), 5157 ms; echo time (TE), 95 ms; field of view (FOV), 30 × 24 cm (phase × frequency); matrix, 128 × 128; slice thickness, 4.0 mm; interslice gap, 0.8 mm; number of slices, 26; number of excitations (NEX), 4; and total acquisition time, 3 min 26 s. DWI was obtained in the axial plane, with the acquisition baseline oriented perpendicular to the long axis of the anal canal.

Image processing and analysis

An abdominal radiologist with 5 years of experience performed manual three-dimensional (3D) whole-lesion segmentation by delineating the abscess contours slice-by-slice on axial DWI images (b = 1000 s/mm²), thereby generating volumetric regions of interest (VOIs). Segmentation aimed to include the entire abscess cavity and wall exhibiting hyperintensity on DWI and corresponding restricted diffusion, while avoiding inclusion of adjacent non-involved tissues. When lesion boundaries were equivocal on DWI, axial T2-weighted fat-suppressed (T2WI-FS) images were used as an anatomic reference to confirm the extent of the abscess. The VOIs were reviewed and confirmed by a senior radiologist with over 15 years of experience in perianal MRI; any disagreements were resolved by joint review to reach consensus. Both radiologists were blinded to all clinical and microbiological information.

To assess interobserver reproducibility, 30 patients included in both Task 1 and Task 2 were selected for independent repeat segmentation by the second reader. Interobserver agreement for abscess volume and the final retained histogram features used to construct the nomograms was assessed using intraclass correlation coefficients (ICCs); features with ICC > 0.75 were considered to have good reproducibility.

DWI images at other b values were co-registered to the b = 1000 s/mm² dataset, and the ROIs were then propagated to the corresponding images to ensure spatial alignment and consistency. Prior to ADC fitting, whole-image DWI datasets at each b value underwent N4 bias field correction and intensity normalization in FireVoxel (version 462) to reduce low-frequency signal inhomogeneity and inter-subject variability. For each b value (400–1600 s/mm²), ADC maps were generated using a monoexponential diffusion model fitted between b = 0 and each higher b value, $S=S_0·\text{exp}\left(-,b,·,A,D,C\right)$, and histogram features were extracted using FireVoxel software (version 462). Quantitative parameters included abscess volume, minimum ADC (minADC), maximum ADC (maxADC), mean ADC, standard deviation, skewness, kurtosis, entropy, and ADC percentiles (1st, 5th, 10th, 25th, 50th, 75th, 90th, 95th, 99th), denoted as P1_ADC, P5_ADC, P10_ADC, and so forth (Fig. 2).

Fig. 2.

Diffusion-weighted imaging (DWI), apparent diffusion coefficient (ADC) maps, and ADC histogram distributions in two patients with perianal abscess. (a-d) Patient with Escherichia coli infection, ESBL-positive. (e–h) Patient with Klebsiella pneumoniae infection, ESBL-negative. (a, e) Axial DWI images; (b, f) Corresponding ADC maps; (c, g) Regions of interest (ROIs) on DWI images; (d, h) Whole-lesion ADC histograms derived from the VOIs

Statistical analysis

Continuous variables were expressed as mean ± standard deviation (SD) or as median with interquartile range, depending on their distribution, and were compared using Student’s t-test or the Mann–Whitney U test. Categorical variables were summarized as counts and percentages and compared using Pearson’s chi-square test or Fisher’s exact test. The Kolmogorov–Smirnov test was used to assess normality. Potential clinical predictors were screened by univariate logistic regression. ADC histogram features were standardized using z-scores. Prior to least absolute shrinkage and selection operator (LASSO), pairwise correlations were evaluated and one variable from each highly correlated pair (correlation coefficient > 0.90) was excluded to reduce collinearity. Imaging features associated with pathogen type or ESBL status were then selected using LASSO regression. The retained imaging features were subsequently linearly combined to generate a composite ADC-derived imaging signature (ADC_Score) using their corresponding coefficients ($ADC\_Score={\beta }_0+{\sum }_{i=1}^n{\beta }_i\times Z\left(F,e,a,t,u,r,e_i\right)$) (coefficients are provided in Supplementary Table S1). The ADC_Score, together with selected clinical variables, was entered into the multivariable logistic regression models to construct the nomograms. A detailed summary of all extracted ADC histogram features across different b values, together with those retained after correlation filtering and LASSO selection for each task, is provided in Supplementary Table S2. Potential interaction effects between clinical predictors and the ADC-derived imaging score were further explored by introducing multiplicative interaction terms into the logistic regression models, with model discrimination compared using area under the curve (AUC) and DeLong’s test. In addition, sensitivity analyses were performed to evaluate potential selection bias related to sex distribution and the availability of ESBL testing. This study was conducted and reported in accordance with the TRIPOD (Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis) guidelines for prediction model development and validation, as shown in the Supplementary Material.

Model performance was evaluated with receiver operating characteristic (ROC) analysis, including calculation of area under the curve (AUC), accuracy, sensitivity, and specificity. Internal validation was performed using bootstrap resampling with 1,000 iterations. Model calibration was assessed with the Hosmer–Lemeshow test and calibration plots, while clinical utility was evaluated using decision curve analysis (DCA). A two-sided p value < 0.05 was considered statistically significant. All statistical analyses were conducted with SPSS (version 27.0) and R software (version 4.5.0). In R, LASSO regression was implemented using the glmnet package; preprocessing was performed using caret; ROC analysis was conducted using pROC; nomograms and calibration were generated using rms; the Hosmer–Lemeshow test was performed using ResourceSelection; and decision curve analysis was conducted using rmda.

Results

Baseline clinical characteristics

Patients were stratified into four groups according to pathogen type and ESBL status. For Task 1 (pathogen identification), patients were classified into the E. coli group (Group 1) or the K. pneumoniae group (Group 2). For Task 2 (ESBL prediction), patients were categorized as ESBL-negative (Group 3) or ESBL-positive (Group 4). Group 1 included 110 patients (99 men, 90%; 11 women, 10%), aged 18–75 years, with a median age of 41 years (interquartile range [IQR], 19.3). Group 2 comprised 47 patients, all male, aged 31–80 years, with a mean age of 46 years (SD, 24). Group 3 consisted of 91 patients (85 men, 93.4%; 6 women, 6.6%), aged 19–80 years, with a median age of 45 years (IQR, 18). Group 4 included 30 patients (27 men, 90%; 3 women, 10%), aged 18–66 years, with a mean age of 40.4 years (SD, 12.7) (Table 1).

Table 1.

Baseline demographic and clinical characteristics of the enrolled population

Characteristics	Pathogen		p	ESBL testing		p
	Group 1 (n = 110)	Group 2 (n = 47)		Group 3 (n = 91)	Group 4 (n = 30)
Age	41.0 (19.3)b	46.0 ± 24.0a	0.008 d	45.0 (18.0)b	40.4 ± 12.7a	0.026 d
Sex			0.035e			0.688e
Male	99 (90.0%)	47 (100.0%)		85 (93.4%)	27 (90.0%)
Female	11 (10.0%)	0		6 (6.6%)	3 (10.0%)
Hypertension			0.014 f			0.633f
Yes	42 (38.2%)	28 (59.6%)		44 (48.4%)	13 (43.3%)
No	68 (61.8%)	19 (40.4%)		47 (51.7%)	17 (56.7%)
Diabetes			< 0.001f			0.108f
Yes	11 (10.0%)	31 (66.0%)		29 (31.9%)	5 (16.7%)
No	99 (90.0%)	16 (34.0%)		62 (68.1%)	25 (83.3%)
Dyslipidemia			0.166f			0.964f
Yes	57 (51.8%)	30 (63.8%)		52 (57.1%)	17 (56.7%)
No	53 (48.2%)	17 (36.2%)		39 (42.9%)	13 (43.3%)
Smoking			0.177f			0.462f
Yes	26 (23.6%)	16 (34.0%)		24 (26.4%)	10 (33.3%)
No	84 (76.4%)	31 (66.0%)		67 (73.6%)	20 (66.7%)
Drinking			0.094f			0.804f
Yes	14 (12.7%)	11 (23.4%)		17 (18.7%)	5 (16.7%)
No	96 (87.3%)	36 (76.6%)		74 (81.3%)	25 (83.3%)
WBC (× 109/L)	11.2 (4.1)b	10.4 ± 3.0a	0.085d	10.6 ± 3.1a	11.7 ± 3.2a	0.085c
Abscess volume (cm3)	10.0 (17.0)b	9.4 (11.0)b	0.130d	9.9 (15.1)b	13.1 (23.6)b	0.287d
Abscess location			0.347e			0.708e
Supralevator	11 (10.0%)	2 (4.3%)		7 (7.7%)	3 (10.0%)
Infralevator	99 (90.0%)	45 (95.7%)		84 (92.3%)	27 (90.0%)
Anal fistula			0.508f			0.878f
Yes	46 (41.8%)	17 (36.2%)		41 (45.1%)	14 (46.7%)
No	64 (58.2%)	30 (63.8%)		50 (55.0%)	16 (53.3%)

^a, normally distributed data (mean ± standard deviation); ^b, non-normally distributed data (median [interquartile range]); other values are presented as number (percentage). Statistical tests: ^c, Student’s t-test; ^d, Mann–Whitney U test; ^e, Fisher’s exact test; ^f, Pearson’s chi-square test. Group 1 represents Escherichia coli; Group 2 represents Klebsiella pneumoniae; Group 3 represents ESBL-negative cases; and Group 4 represents ESBL-positive cases. WBC, white blood cell count

In Task 1, patients in Group 1 were significantly younger (p = 0.008), had a higher proportion of women (p = 0.035), and showed lower rates of hypertension (p = 0.014) and diabetes (p < 0.001) compared with Group 2. In Task 2, patients in Group 3 were older than those in Group 4 (p = 0.026) (Table 1).

Univariate analysis of clinical risk factors

To assess clinical factors associated with pathogen type and ESBL infection in perianal abscesses, univariate logistic regression was conducted to compare Group 1 with Group 2, and Group 3 with Group 4. Variables considered included age, sex, white blood cell (WBC) count, abscess location, hypertension, diabetes, dyslipidemia, smoking, drinking, and the presence of anal fistula. In Task 1, significant associations were identified for age (p = 0.006; 95% confidence interval [CI], 1.01–1.06), hypertension (p = 0.015; 95% CI, 1.19–4.80), and diabetes (p < 0.001; 95% CI, 7.33–41.51), indicating that older age, hypertension, and particularly diabetes were associated with a higher likelihood of K. pneumoniae infection in univariate analysis. In Task 2, age was significantly associated with ESBL status (p = 0.020; 95% CI, 0.93–0.99) (Table 2).

Table 2.

Univariate logistic regression analysis of clinical risk factors for perianal abscess

Characteristics	Pathogen identification		p	ESBL prediction		p
	B	OR (95% CI)		B	OR (95% CI)
Age	0.035	1.04 (1.01–1.06)	0.006	− 0.041	0.96 (0.93–0.99)	0.020
Sex	− 20.458	0.00	0.999	0.454	1.57 (0.37–6.72)	0.540
Hypertension	0.870	2.39 (1.19–4.80)	0.015	− 0.202	0.82 (0.36–1.88)	0.633
Diabetes	2.859	17.44 (7.33–41.51)	< 0.001	− 0.850	0.43 (0.15–1.23)	0.115
Dyslipidemia	0.495	1.64 (0.81–3.31)	0.17	− 0.019	0.98 (0.43–2.26)	0.964
Smoking	0.511	1.67 (0.79–3.52)	0.180	0.333	1.40 (0.57–3.40)	0.463
Drinking	0.740	2.10 (0.87–5.04)	0.099	− 0.139	0.87 (0.29–2.60)	0.804
WBC (× 109/L)	− 0.098	0.91 (0.81–1.02)	0.091	0.116	1.12 (0.98–1.28)	0.089
Abscess location	− 0.916	0.40 (0.09–1.88)	0.246	0.288	1.33 (0.32–5.52)	0.691
Anal fistula	− 0.238	0.79 (0.39–1.60)	0.509	0.065	1.07 (0.47–2.44)	0.878

B, regression coefficient (log odds); OR, odds ratio; CI, confidence interval; WBC, white blood cell count

Feature selection by LASSO regression

From the 97 ADC histogram parameters generated across six b values (400, 600, 800, 1000, 1500, and 1600 s/mm²), variables with correlation coefficients greater than 0.90 were excluded using the findCorrelation function. The remaining features were analyzed using LASSO regression, with the penalty parameter λ determined by tenfold cross-validation (Fig. 3). By applying L1 regularization to the regression coefficients, LASSO enabled automatic selection of variables, effectively reducing multicollinearity and overfitting while enhancing model robustness and predictive performance [24]. A total of 20 features for Task 1 (pathogen discrimination) and 13 features for Task 2 (ESBL prediction) were retained and subsequently combined into composite ADC_Score values (Fig. 3). Logistic regression models based on these composite scores alone (ADC Model 1 for pathogen discrimination and ADC Model 2 for ESBL prediction) achieved AUCs of 0.749 (95% CI, 0.668–0.830) and 0.763 (95% CI, 0.668–0.859), respectively (Fig. 4).

Fig. 3.

Selection of ADC histogram features using the least absolute shrinkage and selection operator (LASSO) regression. (a-c) Task 1: pathogen discrimination. (d-f) Task 2: ESBL infection risk prediction. (a, d) Ten-fold cross-validation curves used to determine the penalty parameter (λ). (b, e) LASSO coefficient profiles illustrating the shrinkage of feature coefficients with increasing λ. (c, f) Selected features with their corresponding regression coefficients. LASSO identified 20 features for pathogen discrimination and 13 features for ESBL risk prediction

Fig. 4.

Receiver operating characteristic (ROC) curves for the predictive models of pathogen discrimination (a-c) and ESBL infection risk prediction (d-f). (a, d) Models constructed using ADC histogram features alone. (b, e) Models constructed using clinical variables alone. (c, f) Nomogram models integrating ADC histogram features with clinical variables

Construction and validation of nomogram models

Multivariable nomograms were developed by integrating ADC_Score with clinical predictors: Nomogram 1 for pathogen discrimination and Nomogram 2 for ESBL prediction (Fig. 5). In addition, interobserver reproducibility in the 30 cases with repeat segmentation was good, with an ICC of 0.895 for abscess volume and ICCs ranging from 0.754 to 0.970 for the retained ADC histogram features used to derive ADC_Score (Supplementary Table S3). Nomogram 1, which incorporated ADC_Score together with age, hypertension, and diabetes, achieved an AUC of 0.897 (95% CI, 0.846–0.949). Nomogram 2, combining ADC_Score with white blood cell count (WBC) and age, yielded an AUC of 0.823 (95% CI, 0.746–0.901) (Fig. 4, Table 3).

Fig. 5.

Nomograms for preoperative prediction of pathogen type and ESBL infection risk in perianal abscess. (a) Pathogen discrimination between Klebsiella pneumoniae and Escherichia coli, incorporating age, diabetes, hypertension, and the ADC-derived composite score (ADC_Score). (b) Prediction of ESBL infection risk, incorporating white blood cell count (WBC), age and ADC_Score. Values assigned to each predictor are scored on a 0–100 scale. By summing the scores for all predictors, a total score is obtained and then mapped to the probability scale to estimate individualized risk

Table 3.

Predictive performances of ADC-only, clinical-only, and combined nomogram models for Task 1 and Task 2

Models	AUC (95% CI)	Accuracy	Sensitivity	Specificity
Task 1
ADC Model 1	0.749 (0.668–0.830)	0.669	0.745	0.636
Clinical Model 1	0.812 (0.738–0.886)	0.841	0.702	0.900
Nomogram 1	0.897 (0.846–0.949)	0.828	0.872	0.809
Task 2
ADC Model 2	0.763 (0.668–0.859)	0.752	0.600	0.802
Clinical Model 2	0.665 (0.556–0.774)	0.488	0.967	0.330
Nomogram 2	0.823 (0.746–0.901)	0.711	0.867	0.659

Task 1, pathogen identification; Task 2, ESBL prediction; CI, confidence interval

Head-to-head comparisons against baseline models are summarized in Table 3. In Task 1, Nomogram 1 significantly outperformed both the clinical-only model (AUC 0.812; DeLong’s p = 0.004) and the ADC-only model (AUC 0.749; DeLong’s p < 0.001). In Task 2, Nomogram 2 showed significantly higher discrimination than the clinical-only model (AUC 0.665; DeLong’s p = 0.002), while its advantage over the ADC-only model did not reach statistical significance (AUC 0.763; DeLong’s p = 0.080).

Exploratory interaction analyses showed that inclusion of interaction terms between clinical predictors and ADC_Score did not significantly improve discrimination for either task (all ΔAUC ≤ 0.006; all DeLong’s p > 0.10; Supplementary Table S4). Calibration analysis demonstrated good agreement between predicted and observed outcomes, with Hosmer–Lemeshow p values of 0.176 for Nomogram 1 and 0.518 for Nomogram 2 (Fig. 6). DCA indicated potential net benefit across clinically relevant threshold probabilities: Nomogram 1 showed higher net benefit over comparator models between 0.05 and 0.75, whereas Nomogram 2 maintained positive net benefit across a broad range of threshold probabilities (approximately 0.05–0.80) (Fig. 6).

Fig. 6.

Calibration curves and decision curve analyses of the nomogram models for pathogen discrimination (a, b) and ESBL infection risk prediction (c, d). (a, c) Calibration curves of the nomogram models, demonstrating consistency between predicted and observed probabilities. (b, d) Decision curve analyses evaluating the net benefit of the nomogram models compared with ADC-only models and clinical-only models, treat-all, and treat-none strategies across varying threshold probabilities

Sensitivity analyses addressing potential selection bias

Additional sensitivity analyses were conducted to evaluate the potential impact of sex imbalance in the K. pneumoniae group and the exclusion of patients without ESBL testing. When analyses were restricted to male patients only, Nomogram 1 remained superior to the ADC-only model, with an AUC of 0.891 (95% CI, 0.835–0.946) compared with 0.762 (95% CI, 0.678–0.846; DeLong’s p < 0.001). Calibration curves demonstrated good agreement between predicted and observed probabilities (Supplementary Fig. S1).

Baseline clinical characteristics were further compared between patients with and without available ESBL testing. No statistically significant differences in baseline characteristics were observed between the two groups (all p > 0.05; Supplementary Table S5), suggesting that no evident differences in measured baseline characteristics were observed between groups, although unmeasured confounding cannot be excluded.

Discussion

In this study, we developed and internally validated two nomograms integrating multi-b-value DWI-derived ADC histogram features with clinical variables, aiming to address the current lack of reliable preoperative approaches for pathogen identification and antibiotic resistance risk stratification in perianal abscesses. The pathogen discrimination model, which distinguished K. pneumoniae from E. coli, achieved an AUC of 0.897, whereas the ESBL prediction model yielded an AUC of 0.823. Both models demonstrated good calibration and offered net clinical benefit on DCA, highlighting the potential of combining ADC histogram features with routine clinical data to enable noninvasive stratification of pathogen type and ESBL production in patients with perianal abscesses. Practically, in this internally validated setting, for patients whose nomogram-predicted probability of ESBL infection exceeds a clinically selected threshold, early ESBL-active empirical coverage may be cautiously considered in selected cases while awaiting culture results, with appropriate antimicrobial stewardship and prompt de-escalation.

From an epidemiological perspective, our study confirmed that perianal abscesses occur predominantly in men, with a mean age of onset of 44.5 years, consistent with previous reports [17, 25]. Further analysis suggested that advanced age, diabetes, and hypertension may serve as risk factors for K. pneumoniae infection. Jeong et al. [26] reported E. coli as the predominant pathogen in non-diabetic patients with perianal abscesses, whereas K. pneumoniae was more frequently isolated in diabetic patients. Similarly, Liu et al. [27] demonstrated that K. pneumoniae was the most common pathogen among diabetic individuals, aligning with our findings. Although few studies have directly compared E. coli and K. pneumoniae infections in perianal abscesses with respect to age and hypertension, existing evidence indicates that diabetes frequently co-occurs with older age and hypertension [28, 29]. The higher prevalence of diabetes in the K. pneumoniae group in our cohort may therefore, at least in part, explain the observed distribution of age and hypertension. Regarding ESBL prediction, age and WBC count were retained as clinical predictors together with ADC_Score. However, although the combined model showed better discrimination than the clinical-only model, its performance was not statistically superior to that of the ADC-only model, indicating that the incremental contribution of clinical variables for ESBL prediction remains limited in the present dataset. This limited clinical contribution is consistent with previous studies: while we observed age-related differences between ESBL-positive and ESBL-negative groups, such associations were not replicated in the studies by Herindrainy [30] and Richelsen [13], which may be attributable to the limited sample size of our study. Although WBC did not reach the conventional significance threshold in multivariable analysis (p = 0.089), it was retained in the ESBL model given its routine availability and biological plausibility as a marker of inflammatory response, supported by prior evidence that ESBL-producing infections can be associated with a more pronounced inflammatory phenotype [31], and because its inclusion was associated with a modest improvement in discrimination in our cohort (AUC increased from 0.815, 95% CI 0.734–0.895 to 0.823, 95% CI 0.746–0.901).

In recent years, DWI has been increasingly applied in the evaluation of perianal abscesses and fistulas, not only for diagnosis but also for assessing disease activity and postoperative outcomes [17, 19, 32, 33]. ADC, an essential quantitative parameter derived from DWI, reflects the degree of water diffusivity within tissues and consequently functions as an indirect marker of pathological changes such as edema, inflammation, and cellular density [34]. Traditional ADC analysis, which typically relies on mean values to represent overall diffusion, often overlooks intralesional heterogeneity [20]. Histogram analysis, by leveraging the distribution of lesion ADC values, enables extraction of statistical parameters such as extreme indices (minimum, maximum), percentiles, distributional parameters (kurtosis, skewness, entropy), and dispersion metrics (standard deviation), thereby providing a more comprehensive assessment of lesion complexity. Although higher-order texture features may provide complementary information on spatial heterogeneity, we focused on first-order whole-lesion ADC histogram descriptors to prioritize interpretability and robustness across imaging protocols. In the present study, we further incorporated multi-b-value DWI to enrich the analytic dimension: high b values primarily reflect molecular self-diffusion and are more sensitive to restricted diffusion, whereas low b values capture perfusion-related components [35]. These complementary measurements provide a fuller description of tissue properties. Moreover, multi-b-value DWI can be integrated with the monoexponential, intravoxel incoherent motion (IVIM), and stretched exponential models (SEM) to separate and quantify free water diffusion, pseudodiffusion, and non-Gaussian diffusion components [36, 37]. Such multiparametric approaches allow a more refined depiction of tissue microstructure and may ultimately facilitate the diagnosis and classification of a range of pathological processes [37].

In this study, the selected multi-b-value DWI histogram features included extreme indices, distributional parameters, dispersion metrics, and lesion volume, underscoring that diverse diffusion-related information contributes meaningfully to both classification and prediction tasks. In Task 1 (pathogen discrimination), features were distributed across b values ranging from 400 to 1600 s/mm²: intermediate b values predominantly provided distributional parameters, low b values contributed extreme indices, and high b values primarily provided percentile and dispersion metrics. In Task 2 (ESBL prediction), LASSO regression likewise identified features spanning low, intermediate, and high b values: kurtosis and entropy were predominant at intermediate-to-high b values, low percentiles were derived from low b values, and high percentiles together with extreme indices were captured at high b values. Overall, the features selected for both tasks encompassed a broad range of diffusion sensitivities, underscoring the strength of multi-b-value histogram analysis in delivering complementary imaging information.

This study combined multi-b-value DWI ADC histogram parameters with clinical features to achieve preoperative prediction of major pathogen types and the risk of ESBL infection in perianal abscesses. Both models demonstrated good discriminative performance, achieving results comparable to those reported for histogram-based analyses in other infectious and oncological conditions [23, 38]. The DCA further revealed consistent net clinical benefit across a wide range of threshold probabilities, underscoring their potential value in preoperative risk stratification rather than direct clinical decision-making. In contrast to the conventional reliance on postoperative pus culture, imaging-based preoperative prediction provides a reproducible and noninvasive alternative. In this internally validated setting, such approaches may help inform empirical antibiotic considerations while awaiting culture results, with appropriate antimicrobial stewardship and prompt de-escalation.

This study had several limitations. First, this was a single-center retrospective study with a relatively modest sample size, and the nomograms were validated using bootstrap resampling only. Although bootstrap validation helps estimate and reduce optimism, the reported performance may still be optimistic in the absence of external or temporal validation. Moreover, given the limited number of outcome events—particularly for ESBL-positive cases—splitting the cohort for temporal validation was unlikely to yield stable estimates. Therefore, external or temporal validation in independent cohorts is warranted before broader clinical use. Second, given the relatively high dimensionality of candidate histogram features and the limited number of events, residual overfitting cannot be fully excluded despite collinearity pruning, LASSO-based feature selection, and bootstrap validation. Third, in Task 1, the K. pneumoniae group comprised only male patients, which may limit the generalizability of the pathogen discrimination model to female patients. To partially address this concern, we performed a sex-restricted sensitivity analysis in male patients only, which yielded consistent discriminative performance, and we also compared baseline clinical characteristics between patients with and without available ESBL testing; these analyses suggested no major imbalances in key clinical variables and supported the robustness of the main findings, although they do not eliminate the need for independent validation. Fourth, our analysis was restricted to E. coli and K. pneumoniae—the two predominant pathogens in perianal abscesses—while other potentially common pathogens were not considered. Although this limits broader applicability, the focus remains clinically meaningful, as these two bacteria are the predominant pathogens and principal ESBL producers [7, 30]. Finally, differences in MRI equipment and acquisition parameters may affect DWI signals and ADC values. Future research should aim to standardize imaging protocols across centers to enhance the robustness and generalizability of the models.

In conclusion, this study developed and validated nomogram models combining multi-b-value DWI-derived ADC histogram features with clinical variables, which exhibited good discriminative performance and promising internal utility for predicting predominant pathogens and assessing the risk of ESBL infections in perianal abscesses. These models provide a noninvasive and quantitative tool for preoperative evaluation of pathogen type and antimicrobial resistance but require external validation before clinical implementation. Looking forward, future studies that incorporate multimodal imaging, radiomics, and deep learning may further enhance predictive accuracy and facilitate broader clinical adoption following independent validation [39].

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

All authors contributed to the study conception and design. Jiajia Wang, Chuanyan Li and Yan Li: Conceptualization, Methodology, Writing-original draft preparation. Ziqi Tang: Investigation, Data curation, Writing-original draft preparation. Na Jiang: Formal analysis. Guangjie Sun: Investigation. Ying Wang, Bingcang Huang and Weiping Lu: Supervision, Writing-review and editing. All authors read and approved the final manuscript.

Funding

This work was supported by The Academic Leaders Training Program of Shanghai Pudong New Area Health Commission (PWRd2024-03), The Investigator-initiated Trial Program of Shanghai Pudong New Area Health Commission (the Medical and Industrial Integration Program, 2026-PWYC-23), The Medical Discipline Construction Program of Shanghai Pudong New Area Health Commission (the Key Disciplines Program, PWZxk2022-03), The Science and Technology Commission of Shanghai Municipality (24SF1904203, 24SF1904200), National Natural Science Foundation of China (82372029).

Data availability

The datasets used or analyzed during the present study are available from the corresponding author on reasonable request.

Declarations

Ethics approval

This study was conducted in accordance with the principles of the Declaration of Helsinki and was approved by the Ethics Committee of Gongli Hospital, Pudong New Area, Shanghai, China (Approval No. 2022–49). Written informed consent was obtained from all participants prior to their inclusion in the study.

Competing interests

The authors declare no competing interests.