Study patientsThis retrospective study was conducted in accordance to the principles stated in the Declaration of Helsinki and approved by the institutional review board of Yonsei Health System (IRB No. 4-2022-1191), and the need to obtain informed consent was waived owing to its retrospective nature. In addition, this study adhered to the CLEAR (CheckList for EvaluAtion of Radiomics) reporting guidelines to improve the credibility, reproducibility, and transparency of the study26.We queried the radiological database for patients who had been diagnosed with periampullary cancer and undergone multiphasic contrast-enhanced MRI with the terms “abscess” or “metastasis” included in radiological MR reports uploaded between October 2009 and December 2019 at hospital 1 (Gangnam Severance Hospital) and between October 2005 and December 2019 at hospital 2 (Severance Hospital). The exclusion criteria for patients were as follows: (1) no focal liver lesions; (2) size of focal hepatic lesions > 2 cm; (3) size of focal hepatic lesions ≤ 5 mm or presenting as tiny diffusion restricting foci without detection on dynamic phases in abdominal MRI; and (4) focal hepatic lesions without follow-up CT or MRI. For the hospital 2 cohort, patients whose fat-saturated T2 weighted images (FS-T2WI) were obtained via echo time (TE) < 80ms and ≥ 85ms were also excluded to closely resemble the TE (83-85ms) of the FS-T2WI protocol of hospital 1. In addition, patients at hospital 2 whose focal liver lesions (1) had typical MRI findings for cystic lesions; (2) had previously received treatment; and (3) were obscured due to MR artifacts were also excluded. Finally, 54 patients with 67 hepatic metastases and 45 hepatic abscesses were allocated from hospital 1 to the training set while 55 patients with 83 hepatic metastases and 40 hepatic abscesses were allocated from hospital 2 to the validation set (Fig. 4). 11 (20.4%) out of 54 patients in the training set and 17 (30.9%) out of 55 patients in the validation set had focal hepatic lesions pathologically confirmed. The rest were confirmed via clinical diagnosis based on follow-up ≥ 6 months with CT or MRI, wherein hepatic abscess was clinically diagnosed if the lesion disappeared or decreased in size with antibiotic treatment at follow-up, and hepatic metastasis was clinically diagnosed if (1) the size of the lesion did not change or increase despite antibiotic treatment, or (2) the lesion persisted, regardless of size increase or decrease, after chemotherapy.Fig. 4Flowchart summarizing patient selection and allocation to the (A) training and (B) validation sets.Sample size calculationIn a previous study, the prevalence of a focal hepatic lesion being metastasis rather than abscess in periampullary cancer patient was 56.9% (41/72)8. To calculate the sample size, we used the method for calculating the sample size required for a clinical prediction model based on the accuracy of estimate27. In brief, the required sample size (n) can be calculating using the anticipated outcome proportion (φ), and the desired margin of error (\(\: \delta\)) using the following equation: \(\:\text{n} = \left( {\frac{{1.96}}{\delta }} \right)^{2} \cdot \phi \left( {1 – \phi } \right).\)Using the above prevalence of 0.569 as the anticipated outcome proportion and setting the margin of error at 0.1, the minimum required sample size was calculated to be 95 patients. In the current study, 109 patients were included, exceeding the minimum sample size, which allows us to develop a reliable prediction model.MRI acquisitionMultiphasic contrast-enhanced MRI was performed with either a 1.5-T or 3.0-T MRI scanner. Routine protocols included T1-weighted three-dimensional gradient-echo imaging with dynamic contrast enhancement, respiratory triggered or breath-hold T2-weighted imaging (WI) with or without fat suppression, and diffusion-weighted imaging (DWI). Details in MR parameters and protocols are provided in Additional File 1, Supplementary Table S4.MR radiomics analysisImage segmentationOne radiologist (J.H.P.) used 3D Slicer version 4.10 (www.slicer.org), a free and open-source software, to semi-automatically segment the entire area of focal liver lesions on three-dimensional, gradient-echo axial T1-WI, T2-WI, and AP images for feature extraction. Segmentation masks from unenhanced T1-WI was registered to T2-WI and AP images. All segmentation masks were confirmed by a senior radiologist (J.C.), and disagreements were resolved in consensus. Both radiologists were blinded to clinical and histopathologic data during this process. Another board-certified radiologist (J.Y.) independently performed tumor segmentation on the training set to analyze interobserver reproducibility.Radiomics feature extractionPrior to feature extraction, the Synthetic Minority Oversampling Technique (SMOTE) algorithm was used to address class imbalance between the number of hepatic metastases and abscesses by generating synthetic data based on the existing five nearest neighbors. Radiomics features were extracted separately for unenhanced T1-WI, T2-WI and AP images using PyRadiomics, an open-source Python package (version 2.1.2; https://pyradiomics.readthedocs.io), via radiomics.featureextractor.RadiomicsFeatureExtractor class28. The following settings were used as arguments: {‘binWidth’: 20, ‘resampledPixelSpacing’: [3, 3, 3], ‘interpolator’: sitk.sitkBSpline, ‘normalize’: True, ‘normalizeScale’: 1, ‘removeOutliers’: True, ‘sigma’: [-3, 3]}. All image types were enabled, including original, wavelet, Laplacian of Gaussian filter, square, square root, logarithm, exponential, gradient, local binary pattern 2D, and local binary pattern 3D. Features were normalized for each MRI sequence using the z-score normalization method. A total of 873 features were extracted for each MRI sequence.Feature selection and building the classification modelOnly radiomics features with good interobserver reproducibility (intraclass correlation coefficient [ICC] > 0.75) were included in the analysis and 692, 682, and 385 features from unenhanced T1-WI, T2-WI and AP images, respectively, in the training set met this criterion. Subsequently, the least absolute shrinkage and selection operator (LASSO) method was used to select the most useful predictive features from the training set. In the LASSO method, 10-fold cross-validation was used to select the optimal regularization parameter alpha, as the average mean squared error of each patient was the smallest. With the optimal alpha, features having a nonzero coefficient in LASSO were considered robust predictors. The support vector machine29 with a linear kernel was used to construct models using features from one of or combinations of unenhanced T1-WI, T2-WI and AP images.ImplementationAll codes were written and run on Google Colab (https://colab.research.google.com, n.d.), which provides 12GB of RAM and an NVIDIA Tesla K80 GPU. Python 3.10.4 was used along with the Python libraries NumPy, Pandas, Scikit-learn, and EDSR. Codes used for radiomics modeling and data analysis have been deposited into a publicly accessible repository (https://github.com/jhp0510/Metastasis-vs.-Abscess—Radiomics-).MRI visual assessmentTwo board-certified abdominal radiologists (H.R. and Y.E.C. with 9 and 16 years of experience, respectively) retrospectively and independently reviewed the images in the validation set. One radiologist (J.H.P) measured the size of each liver lesion and recorded its section number 2 weeks prior to image analysis. Both readers were blinded to the clinical or histopathological results of each case. MRIs of both groups were presented randomly in a blinded manner to avoid bias.For qualitative analysis, the following imaging parameters were evaluated: (a) rim enhancement on each phase including DP for ECA-MRI and TP and HBP for HBA-MRI, (b) low signal intensity (SI) rim on HBP for HBA-MRI and (c) DWI pattern at b value of 800–900 s/mm2, which was subdivided into three categories: (1) homogenous high SI relative to liver parenchyma through the whole area of the focal hepatic lesion, (2) high SI rim confined to the periphery of the focal hepatic lesion with relatively lower SI in the center, and (3) high SI confined only to the center of the focal hepatic lesion. In addition, size discrepancies, which were defined as ≥ 30% difference in the longest diameter of the lesion between unenhanced T1-WI and T2-WI as well as unenhanced T1-WI and HBP were also analyzed. These imaging parameters were selected because previous studies reported them as for differentiating hepatic metastases and abscesses8,16,17,22. Interobserver agreement was assessed for MR imaging parameters after the first independent image analysis. In addition to the above parameters, two readers scored the final interpretation (hepatic metastasis vs. abscess) before and after using the radiomics model. Both readers also evaluated their confidence in diagnoses with a 5-point scoring system (1-least confident; 5-highly confident) before and after using the radiomics model.Study outcomesThe primary outcome was the diagnostic performance, specifically the area under the curve (AUC) of the receiver operating characteristic (ROC) curve, for various MRI-based radiomics models constructed using different combinations of T1-WI, T2-WI, and AP images. The secondary outcomes included the diagnostic performance of the most relevant visually assessable MRI findings from multivariable regression analysis and the performance of radiologists with and without the use of the best-performing MRI-based radiomics model.Statistical analysesThe Shapiro-Wilk test was used to assess normality. Non-normally distributed continuous variables are presented as medians (interquartile range, IQR). Continuous variables were compared using the Mann-Whitney U test while categorical variables were assessed using the X2-test or Fisher’s exact test. The ICC was calculated to evaluate interobserver agreement between radiologists and interobserver reproducibility for radiomics features: an ICC > 0.75 was considered to indicate good reproducibility30. Interobserver agreement between radiologists was expressed by Cohen’s kappa coefficient. The diagnostic performances of the radiomics model and radiologists for predicting hepatic metastases were calculated and the AUCs of receiver operating characteristics curve were compared using the DeLong method. Statistical analyses were performed using R version 3.4.3 (R Foundation for Statistical Computing). Two-sided P < 0.05 was considered statistically significant.
