Radiomics based on diffusion-weighted imaging for differentiation between focal-type autoimmune pancreatitis and pancreatic carcinoma

Abstract

Objective:

To evaluate the parameters of support vector machine (SVM) using imaging data generated from the apparent diffusion coefficient (ADC) to differentiate between focal-type autoimmune pancreatitis (f-AIP) and pancreatic ductal adenocarcinoma (PDAC) when using SVM based on diffusion-weighted imaging.

Methods:

The 2D-ADC_mean and texture parameters (16 texture features × [non-filter+17 filters]) were retrospectively segmented by 2 readers in 28 patients with f-AIP and 77 patients with pathologically proven PDAC. The diagnostic accuracy of the SVM model was evaluated by receiver operating characteristic curve analysis and calculation of the area under the curve (AUC). Interreader reliability was assessed by intraclass correlation coefficient (ICC).

Results:

The 2D-ADC_mean and 3D-ADC_mean were significantly lower in cases of f-AIP (1.10–1.15 × 10^–3 mm²/s and 1.21–1.23× 10^–3 mm²/s, respectively) vs PDAC (1.29–1.33 × 10^–3 mm²/s and 1.41–1.43 × 10^–3 mm²/s, respectively), with excellent and good interreader reliability, respectively (ICC = 0.909 and 0.891, respectively). Among the texture parameters, energy with exponential filtering yielded the highest AUC (Reader 1: 74.7%, Reader 2: 81.5%), with fair interreader reliability (ICC = 0.707). The non-linear SVM, a combination of 2D-ADC_mean, object volume and exponential-energy showed an AUC value of 96.2% in the testing cohorts.

Conclusion:

Our results suggest that non-linear SVM using a combination of 2D-ADC_mean, object volume, and exponential-energy may assist in differentiating f-AIP from PDAC.

Advances in knowledge:

The radiomics based on an apparent diffusion coefficient value may assist in differentiating f-AIP from PDAC.

Introduction

Autoimmune pancreatitis (AIP), one component of a multiorgan disorder known as IgG4-related disease, ^1,2 is a rare type of chronic pancreatitis. AIP is classified morphologically into diffuse type and focal type. Although IgG4-related pancreatitis can be diagnosed serologically, a previous study reported that 37% of 204 patients with AIP had normal serum IgG4 levels. ³ Both clinicians and radiologists find it challenging to differentiate focal-type AIP (f-AIP) from pancreatic ductal adenocarcinoma (PDAC) ⁴ ; both exhibit similar imaging features. It is therefore essential to confirm a pathological diagnosis of f-AIP using fine needle aspiration under endoscopic ultrasonography (EUS-FNA). Definitive diagnosis of PDAC can be achieved by EUS-FNA when a sufficient and accurately obtained biopsy specimen is collected. However, the issues related to FNA such as technical failure and/or failure to obtain sufficient specimens could be an obstacle for a pathological diagnosis. EUS-FNA-based histological diagnosis could be obtained only in 19 of 44 patients (43%) diagnosed with AIP owing to limited biopsy specimens. ⁵ Moreover, differentiation between f-AIP and PDAC could be difficult because of overlapping histological features, such as lymphoplasmacytic acinar inflammation and storiform fibrosis, ⁶ and the patchy distribution of specific histological findings of AIP. ⁵ Consequently, the role of EUS-FNA in distinguishing f-AIP from PDAC remains controversial.

Differentiation between f-AIP and PDAC has been challenging for CT and MRI diagnosis. ^4,7–10 MRI yields better soft tissue contrast than CT and is thus considered superior for demonstrating distinguishing features. ⁷ Especially, the usefulness of the apparent diffusion coefficient (ADC) value obtained from diffusion-weighted imaging (DWI) has been reported for differentiating f- AIP from PDAC. ^8–10

In recent years, radiomics has emerged as a new analytical method for decision support. ¹¹ A new analytical method in radiomics, called texture analysis, can convert the grey-level intensities and pixel positions within an image into high-dimensional mineable features using a large number of automatically extracted data-characterisation algorithms, ^12,13 deriving quantitative features, called ‘texture features’, that provide a measure of intralesional heterogeneity. ¹⁴ Additionally, texture features could be enhanced by using machine learning (ML) analysis to identify the discriminative features for tumour grade differentiation, treatment response, and prognosis prediction. ¹⁵ The support vector machine (SVM) is an ML approach for classifying individuals through pattern recognition analysis and is considered the optimal method for small-sample classification. ¹⁶ The SVM aims to develop a classification model and has recently been used to provide predictive imaging features for specific diseases. ^17,18 Moreover, the SVM provides high accuracy evaluations, even with a limited number of patients. ¹⁶

We hypothesised that SVM using texture parameters generated from ADC values could assist in differentiating f-AIP and PDAC. However, there are few published reports regarding this method. We aimed to evaluate combinations of parameters for differentiation between f-AIP and PDAC when using SVM based on DWI.

Methods and materials

Patients

This study was approved by our institutional review board, which waived the requirement for informed consent because of the retrospective nature of the study. All data were identified retrospectively by searching patient records in our institutional database. The following inclusion criteria were applied: we included patients who were diagnosed with f-AIP based on Asian diagnostic criteria and/or the International Consensus Diagnostic Criteria and/or surgical resection findings between January 2012 and December 2018, ^2,19 as well as patients with pathologically proven PDAC following surgical resection during the same time period. The following exclusion criteria were applied: diagnosis with diffuse-type AIP; f-AIP but no available MRI data before commencing steroid treatment; and PDAC but no available pre-operative MRI data within 6 months before surgery (Figure 1). Here, f-AIP was defined as the presence of a focal mass involving less than half of the total pancreas or multifocal masses on MRI, per the Asian diagnostic criteria for AIP. ^2,7 A combination of clinical symptoms (swelling or masses of single or multiple organs), histopathological findings (extensive infiltration of IgG4+ plasma cells and fibrosis), haematological examination findings (concentrations of serum IgG4 of >135  mg dl⁻¹), imaging modalities (narrowing of ducts and/or enlargement of organs), and response to steroid treatment was used to diagnose AIP. ^2,19 Surgical resections were not performed on all included patients with f-AIP because histopathological findings are not necessarily required for the diagnosis of f-AIP according to diagnostic criteria. ^2,19

Figure 1.

Flow diagram of patient selection in this study. AIP, autoimmune pancreatitis; PDAC, pancreatic ductal adenocarcinoma

Image acquisition

All MRI studies were performed using a 1.5-Tesla MR unit (Magnetom Avanto or Symphony, Siemens Healthcare, Erlangen, Germany) with a 6-channel body matrix coil. Three-dimensional (3D)-fat-suppressed T ₁ weighted gradient recalled echo sequences with Cartesian sampling, named volumetric interpolated breath-hold examination, T ₂ weighted fast-spin-echo images, DWI with low and high b-values (b = 50, 800 s/mm², respectively), and half-Fourier acquisition single-shot turbo spin-echo sequences were obtained. ADC maps were generated automatically on the operating console. A respiratory-triggered 3D-T ₂ weighted magnetic resonance cholangiopancreatography with maximum intensity projection was also generated. Details of the imaging parameters evaluated in this study are summarised in Supplementary Material 1.

Patient and tumour characteristics

Based on the MR images and clinical diagnostic imaging reports, a radiologist with 5 years of MRI experience (Reader 1) retrospectively evaluated the following patient background and MRI binary variables (i.e. presence/absence of morphological findings): age, sex, duct-penetrating sign, capsule-like rim, peripancreatic strands, and multiple main pancreatic duct (MPD) strictures. Reader 1 had checked all MR images and the imaging findings in the diagnostic reports and confirmed that there was no difference between the clinical reports and MRI findings.

Radiomics

Tumour segmentation for texture feature extraction

In this study, imaging data were extracted in two ways by two readers: (1) the conventional method based on region on interest (ROI) setting on a two-dimensional (2D) plane, performed to measure the mean ADC value (2D-ADC_mean) of the lesion and (2) the 3D volume of interest (VOI)-based measurement, performed to calculate the textural parameters of the lesion. Reader 1 retrospectively performed tumour segmentation for all MRIs using all MRI sequences while blinded to clinical and pathological information. When multiple lesions were detected, only the largest lesion was selected for measurement. Each measurement was performed three times, and the average of the three measurements was applied for the statistical analyses. The 2D ROI was set as a 40–50 mm³ region within the lesion. Next, the 3D VOI was generated by the sum of each ROI obtained from slice-by-slice segmentation of the whole tumour (Figure 2). All slices containing areas with measurable ADC values within the lesion were included, and 3D VOI-based texture analysis was applied. All 3D VOI-based measurements were performed once in each case. Both the ROIs and VOIs were set carefully to include the lesion while excluding cysts, pancreatic ducts, the common bile duct, and normal pancreas. All measurements were performed using a dedicated software prototype (syngo.via frontier; SIEMENS Healthcare). To assess the reproducibility of all continuous variables measured by Reader 1, another radiologist with 5 years of MRI experience (Reader 2) retrospectively evaluated all MRIs in the same manner as Reader 1. Reader 2 was blinded to the clinical and pathological information as well as to the results of Reader 1.

Figure 2.

Segmentation of the 3D VOI analysis using a dedicated software prototype (syngo.via frontier; SIEMENS Healthcare). The region of the 3D VOI was set manually to correspond to the lesion on an axial image of the ADC map (yellow areas). 3D VOI, three-dimensional volume of interest; ADC, apparemt diffusion coefficient.

Texture analysis

Texture analysis was performed for data obtained from 3D VOI-based measurements. Before beginning texture analysis, uniform voxel size resampling was initially performed to reslice ADC images into a uniform size. The texture analysis of each VOI comprised only first-order intensity-based features (10th percentile, 90th percentile, maximum, mean, minimum, mean absolute deviation [MAD], median, range, interquartile range, root mean squared, energy, entropy, kurtosis, skewness, uniformity, variance). All segments were processed using square, square-root, logarithm, exponential, wavelet, and log-sigma filtration. Next, 288 texture features were calculated automatically for each case using software (16 texture features × 17 filters). To compute radiomic features, the prototype software uses the publicly available library PyRadiomics. ^14,20,21 Square, square-root, logarithm, and exponential filtration are additionally computed from non-linearly transformed images. Negative values in the original image were made negative again after transformation. All variables were linearly scaled to the original range before being transformed. Descriptions of texture features and filtrations are presented in Supplementary Material 1. ^22–24

Machine learning

We investigated linear and non-linear SVM as ML algorithms. The SVM algorithm assigns a weight to each feature, and this weight vector defines the hyperplane separating the two classes. One interpretation of these weights is as measures of feature strength for informing group membership. ²⁵ A common problem, however, is the instability of this weight vector across different samples, even from the same data set. For the class size bias between AIP and PDAC (AIP < PDAC), we use a weighting ratio of AIP:PDAC = 3:1.

Cross-validation (CV) is an important tool in ML for testing generalisability of a classifier. In k-fold CV, data are split into k parts, k-1 parts are used for training, one part for testing. A separate classifier is trained on each of the k training sets and evaluated on the corresponding test set. Typically, average performance metrics over all k folds are reported together with a range of variation. 10-fold CV is sometimes suggested as a compromise between bias and training sample size. ²⁶ To tune the SVM, 10-fold CV was performed with 10 patients in each group. The SVM classifier was optimised on its two hyperparameters σ and C; σ denotes the inverse variance of the radial basis function kernel and C is the value of the soft margin penalty. The following ranges were tested: C (0.1, 0.2,0.3,0.4,0.5,0.6,0.7,0.8, 0.9, 1.0); σ = (0.1–1.0 in increments of 0.1). ²⁶

Statistical analysis

The χ² or Fisher’s exact test (for low frequencies) was performed for categorical variables and the Mann–Whitney U test for continuous variables. The intraclass correlation coefficient (ICC) was calculated to assess the interreader reliability for continuous variables measured by readers 1 and 2. ICC values of 0.7–0.8, 0.8–0.9, and >0.90 indicated fair, good, and excellent reliability, respectively. Diagnostic performance was evaluated by constructing a receiver operating characteristic (ROC) curve and calculating the area under the curve (AUC). The Youden index was used to determine cut-off values. The variables that met the following criteria were fed into the ML model: variables with significant between-group differences in the univariate analyses of data by Reader 1 and more than fair interreader reliability. All statistical analyses were performed using SPSS v. 25 (IBM Corp., Armonk, NY) and R v. 4.0.2. The ML model was built by R v. 4.0.2. A p-value of <0.05 was considered statistically significant.

Results

This study included 105 patients (28 with f-AIP; 77 with PDAC). The patient demographics and conventional imaging findings are shown in Table 1. Four of the included patients with f-AIP underwent surgical resection. 96% of patients with PDAC (74/77) underwent pre-operative MRI within 3 months before surgery. There was a significant male predominance among patients with f-AIP (p = 0.03). The duct-penetrating sign, capsule-like rim, and multiple MPD strictures were significantly more frequently present in the f-AIP group than in the PDAC group (p < 0.05; Table 1).

Table 1.

Patient demographic and imaging findings

	AIP (n = 28)	PDAC (n = 77)	P value
Age, years (mean ± SD)	63.2 ± 12.9	67.5 ± 9.0	0.059
Sex
Male	23 (82*)	46 (60)	0.018
Imaging findings				Sensitivity (%)	Specificity (%)	Accuracy (%)	AUC (%)
Duct-penetrating sign	13 (46)	60 (78)	0.003	53.6	76.9	70.8	65.7
Capsule-like rim	17 (61)	3 (4)	< 0.001	60.7	96.1	86.7	78.4
Peripancreatic strands	6 (21)	21 (27)	0.545	28.2	77.8	41.0
Multiple MPD strictures	8 (29)	8 (10)	0.022	22.5	50.0	26.7	59.1

AIP, autoimmune pancreatitis; AUC, area under the curve; MPD, main pancreatic duct; PDAC, pancreatic ductal adenocarcinoma; SD, standard deviation; For univariate analysis,.

^aValues in parentheses indicate percentages.

The 2D-ADC_mean was significantly lower in the f-AIP group than in the PDAC group ( Table 2). The sensitivity, specificity and accuracy of the 2D-ADC_mean were 75.0%, 70.1% and 71.4%, respectively. The AUC values of the 2D-ADC_mean were 84.5% (Reader 1) and 72.6% (Reader 2), and the interreader reliability was excellent (Table 2). The object volume was significantly larger in f-AIP cases than in PDAC cases (p < 0.001). The mean ADC (3D-ADC_mean) was significantly lower in f-AIP cases, and a good level of interreader reliability was observed (p < 0.001; Table 2). Among the unfiltered features, the energy, variance, MAD, and interquartile range showed significant between-group differences for both readers (all p < 0.02; Table 2). Additionally, the kurtosis obtained by Reader 2’s VOIs showed significant between-group differences (p = 0.049; Table 2). Energy had fair interreader reliability; the other features had poor reliability (Table 2).

Table 2.

Results of imaging data obtained from segmentation by two readers

	Reader 1				Reader 2				ICC
	AIP	PDAC	p- value	AUC (%)	AIP	PDAC	p- value	AUC (%)
2D-ADCmean	1.10 ± 0.15	1.33 ± 0.17	<0.001	84.5	1.15 ± 0.18	1.29 ± 0.19	0.001	72.6	0.909
3D-ADCmean	1.23 ± 0.20	1.43 ± 0.21	<0.001	75.9	1.21 ± 0.19	1.41 ± 0.20	<0.001	77.6	0.891
Object volume	16.65 ± 12.64	7.19 ± 5.98	<0.001	74.3	16.23 ± 9.75	7.84 ± 5.341	<0.001	77.4	0.802
Energy (×102)	4.61 ± 4.43	2.42 ± 3.17	0.004	68.2	5.47 ± 5.23	2.08 ± 2.18	<0.001	76.9	0.707
Variance	0.15 ± 0.07	0.20 ± 0.09	0.008	67.0	0.13 ± 0.06	0.19 ± 0.11	0.007	67.3	0.516
Mean absolute deviation	0.31 ± 0.09	0.36 ± 0.09	0.009	66.8	0.28 ± 0.07	0.35 ± 0.10	0.006	67.5	0.583
Interquartile range	0.54 ± 0.19	0.64 ± 0.20	0.016	65.4	0.48 ± 0.14	0.62 ± 0.24	0.007	67.3	0.522

ADC, apparent diffusion coefficient; AIP, autoimmune pancreatitis; AUC, area under the curve; Abbreviations:;ICC, intraclass correlation coefficient; PDAC, pancreatic ductal adenocarcinoma.

The univariate analysis was performed using the Mann–Whitney U test. For variables with statistically significant results, receiver operating characteristic curves were constructed and the AUCs were calculated. All variables showed significant between-group differences in the univariate analyses of data by Reader 1 and more than fair inter reader reliability. Texture parameters were calculated without filtering. Only the features that showed more than moderate AUC (>70%) for either reader are listed, and the full version can be found in the Supplemental.

In a univariate analysis of filtered texture features, 128 and 134 features showed significant differences between f-AIP and PDAC for readers 1 and 2, respectively (all p < 0.05). The list of energy features with and without filtering is presented in Table 3. Of all texture features with and without filtering, energy with exponential filtration (termed exponential-energy henceforth) showed the highest AUC for both readers (Table 3). Exponential-energy was significantly higher in the f-AIP group than in the PDAC group (p < 0.001). Typical cases in this study are shown in Figures 3 and 4.

Table 3.

Results of texture analysis of energy with filtering

	Reader 1				Reader 2
	AIP	PDAC	p- value	AUC (%)	AIP	PDAC	p- value	AUC (%)
Energy (×102)	4.61 ± 4.43	2.42 ± 3.17	0.004	68.2	5.47 ± 5.23	2.08 ± 2.18	<0.001	76.9
Square root-energy (×103)	9.55 ± 8.16	4.45 ± 5.65	<0.001	72.4	11.46 ± 8.41	4.78 ± 8.51	<0.001	78.6
Logarithm-energy (×103)	6.45 ± 6.03	3.21 ± 4.33	0.002	69.9	7.79 ± 6.86	3.35 ± 6.12	<0.001	75.6
Exponential-energy (×103)	16.03 ± 13.83	6.91 ± 5.07	<0.001	74.7	18.20 ± 10.98	6.81 ± 5.51	<0.001	81.5
Wavelet LLL-energy (×103)	36.43 ± 35.14	19.06 ± 25.09	0.004	68.3	43.30 ± 41.47	16.18 ± 17.08	<0.001	74.0
Log sigma 3.5mm-energy (×102)	8.72 ± 6.85	5.71 ± 5.91	0.030	63.9	10.98 ± 7.70	6.24 ± 4.13	0.003	70.3
Log sigma 4.5mm-energy (×102)	10.58 ± 9.45	6.18 ± 6.97	0.022	64.7	13.08 ± 7.70	6.25 ± 4.13	0.001	72.2

ADC, apparent diffusion coefficient; AIP, autoimmune pancreatitis; AUC, area under the curve; PDAC, pancreatic ductal adenocarcinoma.

The univariate analysis was performed using the Mann–Whitney U test. For variables with statistically significant results, receiver operating characteristic curves were constructed and the AUCs were calculated. Only the features that showed more than moderate AUC (>70%) for either reader are listed and the full version can be found in the Supplemental.

Figure 3.

Images of an 84-year-old female with right upper abdominal pain and a loss of appetite. (A, B) A hypointense lesion (white arrows) was detected in the pancreatic head on a fat-suppressed T ₁ weighted image and a T ₂ weighted image. (C, D) A DWI and ADC-map show that the mass (white arrows) exhibited a lower ADC value than the normal pancreatic tissue. The mean ADC values on 2D and 3D analyses and the energy value were 1354.04, 1481.77, and 64.77, respectively. 2D, two-dimensional; 3D, three-dimensional; ADC apparent diffusion coefficient; DWI, diffusion-weighted imaging.

Figure 4.

Images of a 62-year-old male with a mass lesion in the pancreatic head. The mass was detected by screening ultrasonography during an annual medical check-up. (A, B) A hypointense lesion (white arrows) was detected in the pancreatic head on a fat-suppressed T ₁ weighted image and T ₂ weighted image. (C, D) A DWI and ADC-map show that the mass (white arrows) exhibited a lower ADC value than the normal pancreatic tissue. The mean ADC values in the 2D and 3D analyses and the energy value were 985.82, 1177.90, and 4376.61, respectively. 2D, two-dimensional; 3D, three-dimensional; ADC apparent diffusion coefficient; DWI, diffusion-weighted imaging.

For ML, the training cohort included 84 patients and the testing cohort included 21 patients. Their characteristics are shown in Table 4. Age, sex, and AIP percentage were set such that the training and testing cohorts had no significant difference between the training and testing cohorts. The diagnostic performance of linear and non-linear SVM using extracted imaging data for differentiation between f-AIP and PDAC are shown in Table 5. Using linear SVM, a combination of 2D-ADC_mean, 3D-ADC_mean and exponential-energy showed an AUC value of 74.4% in the testing cohorts. Using non-linear SVM, a combination of 2D-ADC_mean, object volume and exponential-energy showed an AUC value of 96.2% in the testing cohorts.

Table 4.

Clinical characteristics of patients in the training and testing cohorts

	Training cohort (n = 84)	Testing cohort (n = 21)	p-value
Age, years (mean ± SD)	66.5 ± 9.9	66.1 ± 12.2	0.885
Sex
Male	55 (65*)	14 (67)	0.918
Label
AIP	22 (26)	6 (29)	0.825

AIP, autoimmune pancreatitis; SD, standard deviation

Table 5.

Diagnostic performance of SVM model using extracted imaging data for differentiation between f-AIP and PDAC

SVM model	Linear		Non-linear
	AUC
Combination	Training cohort (n = 84)	Testing cohort (n = 21)	Training cohort (n = 84)	Testing cohort (n = 21)
2D-ADCmean	87.6	68.9	83.8	81.2
Object volume	74.3	73.3	69.2	82.5
3D-ADCmean	78.3	68.9	74.1	51.2
Exponential－energy	75.6	73.3	70.0	82.5
2D-ADCmean, Object volume	89.0	70.0	82.3	96.2
2D-ADCmean, 3D-ADCmean	86.9	72.2	82.3	81.2
2D-ADCmean, Exponential-energy	75.0	73.3	82.9	93.8
Object volume, 3D-ADCmean	87.9	73.3	84.1	83.8
Object volume, Exponential-energy	90.0	72.2	69.4	87.5
3D-ADCmean, Exponential-energy	87.2	73.3	82.1	81.3
2D-ADCmean, Object volume, 3D-ADCmean	91.0	73.3	83.4	91.3
2D-ADCmean, Object volume, Exponential-energy	90.1	72.2	82.1	96.2
2D-ADCmean, 3D-ADCmean, Exponential-energy	89.4	74.4	83.5	90.0
Object volume, 3D-ADCmean, Exponential-energy	86.6	70.0	82.9	80.0
2D-ADCmean, Object volume, 3D-ADCmean, Exponential-energy	90.5	73.3	83.0	91.3

AUC, the area under the curve; PDAC, pancreatic ductal adenocarcinoma; SVM, support vector machine; f-AIP, focal-type autoimmune pancreatitis.

^aSelection of features are represented in the order Object volume, 2D-ADC_mean, 3D-ADC_mean, Exponential-energy from left to right, and defined as one if the item is included in the model and 0 if not.

Discussion

In this study, we evaluated which combinations of parameters were useful for differentiation between f-AIP and PDAC when using SVM based on imaging data generated from DWI. All ADC-based imaging data were obtained from segmentation by two observers, and VOI data were applied to texture analysis to accurately quantify the intralesional properties. The imaging data that showed high ICC between the two readers and significant differences between f-AIP and PDAC were fed into the SVM to identify useful combinations of parameters to differentiate f-AIP and PDAC.

Among the patient demographic and conventional imaging findings, the sex ratio and presence of the duct-penetrating sign, capsule-like rim, and multiple MPD strictures differed significantly between the f-AIP and PDAC groups; each of these findings yielded lower sensitivity and smaller AUC (<70%) than those associated with texture analyses. Since these findings were reported as useful for diagnosing diffuse-type AIP, ^4,7–9 our results suggest that a detailed evaluation to detect such morphological findings should be performed if a focal pancreatic lesion suspicious for f-AIP or PDAC is detected on MRI.

Both the 2D-ADC_mean and 3D-ADC_mean were significantly lower in f-AIP cases than in PDAC cases. Among texture parameters, exponential-energy yielded the highest AUC for both readers, with good interreader reliability. Both exponential-energy and 3D-ADC_mean were identified as significant factors in the differentiation of f-AIP from PDAC. The 2D-ADC_mean was significantly lower in patients with f-AIP than in those with PDAC. The 2D-ADC_mean could identify f-AIP more accurately than any other morphological finding other than a capsule-like rim. Moreover, the 2D-ADC_mean showed a higher ICC than the texture parameters of the 3D VOI-based analysis. Choi et al reported an AUC of 0.750 (95% CI, 60.75–86.93) for the mean ADC value and an optimal ADC cut-off value of 0.94 × 10^–3 mm²/s for differentiating mass-forming AIP from PDAC. ⁸ In our study, the AUC of the 2D-ADC_mean was 0.73–0.85, and the optimal cut-off value was 1.14–1.19 × 10^–3 mm²/s. Although ADC values depend on the MRI device or parameters (e.g. magnetic field strengths or b-values) used at each institution, our results demonstrate that mean ADC value assessment could feasibly be used to differentiate between f-AIP and PDAC, as reported by Choi et al. ⁸ In particular, the non-linear SVM model using the 2D-ADCmean and object volume demonstrated a diagnostic performance equivalent to those using these two parameters, plus exponential energy, for a total of three parameters. Utilisation of the 2D-ADCmean (which can be easily measured) and object volume (which is a relatively simple parameter among the vast number of texture parameters combined in non-linear SVM models) could achieve the highest diagnostic performance; thus, we believe that this result is feasible to incorporate into clinical practice.

In our texture analysis without filtering, both readers identified energy, variance, MAD, and interquartile range as features that differed significantly between f-AIP and PDAC. Energy, the square sum of each matrix element, yielded the highest AUC with fair interreader reliability. This feature reflects the homogeneity of the greyscale distribution in an image and textural coarseness. ²⁰ High energy might be expected if co-occurrence matrix values are unequal. Recent studies have reported that malignant tumours, which are highly heterogeneous, exhibit higher energy than benign tumours. ^14,27,28 Our study showed that higher energy was a significant factor for differentiating f-AIP from PDAC, and PDAC had a relatively lower energy. Interestingly, one MRI texture analysis study demonstrated that active inflammation increased heterogeneity in a mouse model. ²⁹ However, this has not yet been proven in humans. Exponential filtration takes the exponential function of the absolute intensity linearly scaled to the original range. Although the true effect and interpretation of this filter on texture analysis is unclear, it removes image heterogeneity secondary to noise and highlights biologically important heterogeneity. ²⁹ Although previous studies have evaluated the imaging factors that differentiate f-AIP from PDAC using multivariate analysis, ^8,30 unlike logistic regression, SVMs can prevent the model from being sensitive to data outliers, and cope with non-linearity in the data by using non-linear kernel functions instead of a simple linear kernel. An SVM can thus make good predictions for prospective analyses. ³¹

Our study showed that 2D-ADC_mean, objective volume, 3D-ADC_mean, and exponential-energy were found to be predictive parameters. Among these parameters, non-linear SVM with a combination of 2D-ADC_mean, object volume, and exponential-energy showed a high AUC in the testing cohort. In contrast, SVM with only 2D-ADC_mean or 3D-ADC_mean showed a low AUC in the testing cohort. Thus, our study suggests that a combination of the mean ADC and texture analysis is a feasible and reproducible method for differentiating f-AIP from PDAC.

In this study, especially with linear SVM, the results of the testing cohorts’ performance tended to be lower than those of the training cohorts. These results are presumed to have been caused by overfitting the training cohorts, leading to poor generalisation to previously unknown cohorts. Moreover, misclassification due to data noise and/or data overlap made differentiation between f-AIP and PDAC difficult. Additionally, the results of diagnostic performance were better with non-linear than with linear SVM. Further prospective trials to test the two best models in an unseen cohort of patients with DWI-MR are required to understand the robustness of the texture metric model among different readers. Multicentre, multivendor trials are also needed to investigate the robustness of the texture metric model across different DWI pulse sequences and ADC calculations.

This study had several limitations of note. First, this was a single-centre retrospective study with a limited number of patients. Second, we only analysed first-order features with non-filter and 17 filters. It may be necessary to verify whether additional texture features (e.g. second and third orders) improve the diagnostic performance. Third, although the inflammatory activity in f-AIP might affect the ADC value and lesion heterogeneity, our study failed to provide evidence of inflammatory activity in the affected patients. Further studies are needed to verify the association between the ADC value, heterogeneity, and degree of inflammatory activity in patients with f-AIP.

In conclusion, our results suggest that the non-linear SVM, a combination of 2D-ADC_mean, object volume, and exponential-energy may assist in differentiating f-AIP from PDAC.