Accuracy of diffusion-weighted imaging-magnetic resonance in differentiating functional from non-functional pituitary macro-adenoma and classification of tumor consistency

Abstract

Purpose

The purpose of this study was to determine the accuracy of selected first or second-order histogram features in differentiation of functional types of pituitary macro-adenomas.

Materials and methods

Diffusion-weighted imaging magnetic resonance imaging was performed on 32 patients (age mean±standard deviation = 43.09 ± 11.02 years; min = 22 and max = 65 years) with pituitary macro-adenoma (10 with functional and 22 with non-functional tumors). Histograms of apparent diffusion coefficient were generated from regions of interest and selected first or second-order histogram features were extracted. Collagen contents of the surgically resected tumors were examined histochemically using Masson trichromatic staining and graded as containing <1%, 1–3%, and >3% of collagen.

Results

Among selected first or second-order histogram features, uniformity (p = 0.02), 75th percentile (p = 0.03), and tumor smoothness (p = 0.02) were significantly different between functional and non-functional tumors. Tumor smoothness > 5.7 × 10⁻⁹ (area under the curve = 0.75; 0.56–0.89) had 70% (95% confidence interval = 34.8–93.3%) sensitivity and 33.33% (95% confidence interval = 14.6–57.0%) specificity for diagnosis of functional tumors. Uniformity ≤179.271 had a sensitivity of 60% (95% confidence interval = 26.2–87.8%) and specificity of 90.48% (95% confidence interval = 69.6–98.8%) with area under the curve = 0.76; 0.57–0.89. The 75th percentile >0.7 had a sensitivity of 80% (95% confidence interval = 44.4–97.5%) and specificity of 66.67% (95% confidence interval = 43.0–85.4%) for categorizing tumors to functional and non-functional types (area under the curve = 0.74; 0.55–0.88). Using these cut-offs, smoothness and uniformity are suggested as negative predictive indices (non-functional tumors) whereas 75th percentile is more applicable for diagnosis of functional tumors.

Conclusion

First or second-order histogram features could be helpful in differentiating functional vs non-functional pituitary macro-adenoma tumors.

Introduction

There are several methods for treatment of macro-adenoma including endoscopic trans-sphenoidal hypophysectomy,¹ chemotherapy,² radiotherapy,^3–5 and recently endovascular therapy.^6,7 Surgical resection is an effective operation to cure pituitary adenomas. Nevertheless, tumor extension, consistency, and functionality may determine the rate of successful outcome.^8–11 The consistency of lesion is determinative for neurosurgeons to decide which operation is needed, the transsphenoidal endoscopic surgery or craniotomy.¹² Naturally, hard consistent tumors are elected for craniotomy whereas soft types should be operated with transsphenoidal endoscopy.^12,13 Then, before starting the operation, neurosurgeons need to know the consistency of the brain lesion. Radiologic evaluations have been used preoperatively for diagnosis and assessment of pituitary lesions, such as macro-adenoma.^1,8 Minimally invasive methods are of interest for surgical or radiological interventions, anyway with regards to desired clinical outcome. Most novel therapeutic methods rely on para-clinical evaluations, before starting clinical intervention. A novel magnetic resonance imaging (MRI) technique, diffusion-weighted imaging (DWI), is suspected to show the tumor's hardness using information from water movement during the imaging process.^14–16 A more practical calculated index is the apparent diffusion coefficient (ADC) plus the first and the second-order histogram (FSOH) statistics which are extracted from the region of interest (ROI) when taking diffusion-weighted images.^17,18 There are studies that confirm the applicability of the MRI technique for differential diagnosis of brain tumor types. However, diagnostic test criteria such as accuracy, sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV) for probable cut-offs of FSOH, have not been well discussed and evaluated in the differentiation of functional tumors from non-functional types with regard to clinical application. The aim of the current study is to assess the diagnostic test criteria of selected FSOH statistics in differential diagnosis of functional from non-functional macro-adenomas. Further, we have proposed probable cut-off points for significant FSOHs that may be validated in more comprehensive clinical experiences.

Materials and methods

Study population

This study was approved by the Institutional Review Board. The study population included patients who were referred to the University Hospital from April 2015–April 2016. All included patients who provided their informed written consent underwent MRI, and clinical and radiological assessments indicated diagnosis of macro-adenoma for them. Thirty-two patients were elected for surgical resection; then, their samples were sent to the pathology ward for histochemical staining. Discrimination between functional and non-functional tumors was carried out by previous serum hormone determination.

MRI

All 32 patients underwent MRI on a 1.5 Tesla MRI scanner (Siemens, Avanto, Rel 16.0) using the standard head coil. Magnetic resonance (MR) images were obtained according to the following tumor protocol: axial and sagittal T1 weighted spin echo (SE) (time of repetition/time of echo or (TR/TE) = 400/12 ms), axial and coronal SE T2 weighted images (TR/TE = 3600/97 ms) and axial fluid attenuated inversion recovery images (TR/TE = 4000/117 ms). DWI-EP (echo-planar) sequence was b-values = 0, 800, 1000. Diffusion-weighted images were performed in the axial plane before the contrast administration, using single-shot multi-slice SE (T2 weighted) echo-planar sequence in three orthogonal directions and ADC maps were generated using b-values of 0, 800 and 1000 S/mm². The following features were obtained in all sequences: a matrix size of 384 × 384, section thickness 5 mm, an inter-slice gap of 1 mm, and field of view of 20–24 cm. Post-contrast axial and coronal T1 weighted sequence was performed after intravenous administration of 0.1 mmol gadopentetate dimeglumine (GAD) per kilogram of body weight (Table 1).

Table 1.

Magnetic resonance imaging parameters; tumor protocol was used for imaging with a 1.5 Tesla magnetic resonance imaging (MRI) scanner (Siemens, Avanto, Rel 16.0).

Brain MRI (tumor protocol)	Axial T1	Axial T2	Axial FLAIR	Axial DWI (b value:0–1000)	Post-contrast T1 axial	Post-contrast T1 sagittal	Post-contrast T1 coronal
TR (ms)	428	5791	7510	3600	428	428	428
TE (ms)	9	85	115	102	9	9	9
Flip angle (˚)	90	90	90	-	90	90	90
Field of view (mm2)	205 × 186	210 × 210	205 × 192	230 × 230	205 × 186	210 × 196	200 × 165
Slice thickness (mm)	5	5	5	5	5	5	5
Gap (mm)	1.5	1.5	1.5	1.5	1.5	1.5	1.5
Matrix	256 × 256	261 × 384	224 × 256	192 × 192	256 × 256	255 × 320	224 × 320

ms: millisecond; mm²: square millimeter; mm: millimeter.

Image analysis

All images were transferred to an offline workstation and reviewed independently by a radiologist with four years of experience in brain MRI who was blinded to the clinical and histopathological information of the patients.

Volumetric polygonal whole-lesion borders of the macroadenoma were delineated manually on the whole volumetric slices of contrast-enhanced T1-weighted images by the same radiologist. To avoid partial volume effects, the borders were selected immediately inside the outer margin of the lesion. Additionally, peritumoral edema and hemorrhagic regions of macroadenoma were avoided in determination of the lesion borders, by using T2 and T1 weighted sequences as a guide. Tumor delineation was performed by ImageJ offline software (http://rsb.info.nih.gov/ij/).

ADC maps were automatically co-registered to the counterpart to post-contrast T1-weighted (T1 + GAD) images by rigid intra-subject registration using normalize mutual information (NMI) similarity measure and trilinear transformation method (FSL software, https://fsl.fmrib.ox.ac.uk/).¹⁹ Figure 1 illustrates the overall analysis procedure.

Figure 1.

Schematic flow diagram of the overall classification framework and the method of region of interest (ROI) selection are represented.

Regions of interest (ROIs) of tumorous regions were overlaid on the ADC maps and FSOH analysis was applied. From FSOH analysis, several statistical features including mean ADC, normalized ADC mean (NADC), standard deviation (SD), variance (as a representative of smoothness), skewness, energy (as a representative of uniformity), entropy, kurtosis and the 25th, 50th, and 75th percentiles of the histogram were extracted. NADC was calculated as the ratio of mean ADC of the tumor ROI to that of the normal-appearing white matter. SD shows the contrast of ROI and smoothness represents how gray level changes in ROI. Note that there were other histogram features that statistically, diagnostically, or clinically were not suitable for evaluation, at least in the current work with a limited sample size due to the lack of cases with intended status or acceptance to be included in the study. Thus, only selected FSOH features are evaluated and presented in the current work, to prevent dimensionality bias (when the number of extracted features are more than or near to the number of samples).

Histochemical analysis

The collagen content of tumor specimens was examined histochemically using Masson trichromatic staining quantitatively, for semi-quantitative analyses of collagen content; microscopic photographs were captured in three random regions of each specimen. The MIPAR image processing software program (http://www.mipar.us/) was used for calculating the area of the blue-stained collagen fibers. Collagen contents were graded as <1%, 1–3% and >3% after surgical resection and histological evaluation for each patient.

Statistical analyses

Normal distributions of data were evaluated by the Shapiro-Wilk test which revealed that the distribution of data was not in Gaussian shape; then, the Mann-Whitney U-Test was performed for comparing functional and non-functional tumors. Furthermore, the Kruskal-Wallis test was applied for comparison of groups with collagen contents <1% or 1–3% and >3%. Values of p < 0.05 were considered statistically significant. Receiver-operator characteristic (ROC) curve analysis was performed to obtain the cutoff points for the statistically significant variables, and to determine the sensitivity, specificity, and the area under the curve (AUC) metrics.

Further, to evaluate the positive and negative predictive values (PPV and NPV), we considered the cut-off numbers, obtained from ROC analysis, to categorize tumors as active or inactive according to the smoothness, uniformity and 75th percentile. PPV and NPV were calculated using the formula below and 2 × 2 tables for each significant variables of FSOH and functionality groups (functional and non-functional tumors) according to patient's hormonal status history. PPV and NPV were calculated using MedCalc Statistical Software version 15.8; for this purpose, firstly, we used the associated criterion base on ROC analysis result, for categorizing samples into active (disease present) or inactive (disease absent) tumors. Then, 2 × 2 tables were designated for functionality against the activity of disease.

$$\mathit{PPV}=\frac{\mathit{sensitivity}\times \mathit{prevalence}}{\{\begin{array}{c}\hfill \mathit{sensitivity}\times \mathit{prevalence}\hfill \\ +(1-\mathit{specificity})\times (1-\mathit{prevalence})\end{array}\}}$$

$$\mathit{NPV}=\frac{\mathit{specificity}\times (1-\mathit{prevalence})}{\{\begin{array}{c}\hfill (1-\mathit{sensitivity})\times \mathit{prevalence}\hfill \\ +\mathit{specificity}\times (1-\mathit{prevalence})\end{array}\}}$$

Results

MR image acquisition and histopathological assessment were performed for 32 patients; 21 (65.6%) males and 11 (34.4%) females. The population age mean±SD was 43.09 ± 11.02 (min = 22 and max = 65 years); the mean±SD of male and female patients were 43.61 ± 11.13 and 42.09 ± 11.25, respectively and not different statistically (p = 0.716). Semi-quantitative assessment of collagen contents of resected tissues performed using histochemical staining method; the collagen contents were ≤1% in 16 (50%), 1–3% in 12 (37.5%) and ≥3% in four (12.5%) resected and pathologically diagnosed samples. Furthermore, 10 patients (31%) were diagnosed to be involved with functional and 22 (69%) with non-functional tumors. Figures 2 and 3 show various hard and soft tumor examples.

Figure 2.

Pituitary adenoma with hard consistency. (a) contrast-enhanced T1-weighted image (CE-T1-WI) shows heterogeneous enhanced pituitary mass with intra-tumoral intense dots; (b) T2-WI shows an isointense mass with respect to normal white matter with intra-tumoral bright dots; (c) Histologic examination of resected tumors shows neoplastic tissue composed of large group of small monomorphic cells surrounded by collagen fibers which are represented as bluish color parts using Masson trichrome staining method.

Figure 3.

Pituitary adenoma with soft consistency; (a) contrast-enhanced T1-weighted image (CE-T1-WI) shows large homogenous enhanced pituitary mass with extension into cavernous sinuses; (b) T2-WI shows a relatively hyper intense mass with respect to normal white matter with extension into cavernous sinuses; (c) histologic examination of the resected tumor shows neoplastic tissue composed of large group of small monomorphic cells with scant collagen fibers which are bluish in color (Masson trichrome staining method).

Table 2 shows comparison of the mean values of FSOH features among functional and non-functional tumors. The statistically significant parameters for differentiation of functional and non-functional tumors comprised of tumor smoothness (p = 0.025), uniformity (p = 0.022) and 75th percentile (p = 0.035). FSOH features were not meaningfully different in ADC values when considering the qualitative collagen contents of the tumors (Figure 4 and Table 3). According to Table 4, smoothness with values higher than 5.7058 × 10⁻⁹ has at least 70% (95% confidence interval (CI) = 34.8–93.3%) sensitivity and 33.33% (95% CI = 14.6–57.0%) specificity for diagnosis of functional tumors (ROC p-value = 0.0134) with AUC = 0.752 and standard error (SE) = 0.102 (95% CI for SE = 0.565–0.889). Uniformity ≤ 179.271 had about 60% (95% CI = 26.2–87.8%) sensitivity and 90.48% (95% CI = 69.6–98.8%) specificity for diagnosis of functional tumors from non-functional types (ROC p-value = 0.0082) with AUC = 0.757 and SE = 0.0973 (95% CI = 0.570–0.892). Additionally, 75th percentile with values higher than 0.7009 showed 80% (95% CI = 44.4–97.5%) sensitivity and 66.67% (95% CI = 43.0–85.4%) specificity for categorizing functional and non-functional tumors (ROC p value = 0.0088) with AUC = 0.738 and SE = 0.0909 (95% CI = 0.550–0.879). Figures 5 and 6 depict cumulative ROC curves for statistically significant variables.

Table 2.

Comparison of apparent diffusion coefficient (ADC) values obtained from magnetic resonance (MR) imaging between functional and non-functional tumors; as it is obvious, by categorization of tumors to functional (Yes) and non-functional (No) types, at least smoothness, uniformity and 75th percentile may be useful for differentiation of tumors.

FSOH	Functional	ADC mean±SD	ADC median (min, max)	p-Value
Maximum	No	2674.44 ± 410.764	2675.2 (2045, 3425)	0.272
	Yes	2391.655 ± 588.971	2579.4 (1353, 3220.333)
Mean	No	1169.985 ± 367.412	1098.813 (627.214, 1864.688)	0.735
	Yes	1065.155 ± 385.707	1075.668 (423.46, 1879.57)
Median	No	1078.929 ± 407.863	983 (557.5, 1876.5)	0.473
	Yes	971.5 ± 391.206	884 (485, 1855.5)
Minimum	No	387.443 ± 220.369	391.75 (76, 1023)	0.291
	Yes	399.718 ± 449.767	262.083 (0, 1557)
Normal mean	No	0.381 ± 0.125	0.339 (0.237, 0.751)	0.612
	Yes	0.409 ± 0.134	0.387 (0.28, 0.718)
Standard deviation	No	0.156 ± 0.039	0.153 (0.092, 0.246)	0.091
	Yes	0.192 ± 0.067	0.183 (0.07, 0.295)
Smoothness	No Yes	5.66 × 10−8±1.68 × 10−7 1.51 × 10−7±2.27 × 10−7	2.85 × 10−9 (5.59 × 10−12, 7.22 × 10−7) 4.70 × 10−8 (3.00 × 10−10, 6.26 × 10−7)	0.025a
Third moment	No	0.003 ± 0.003	0.003 (–0.003, 0.007)	0.310
	Yes	0.006 ± 0.006	0.004 (0, 0.019)
Uniformity	No	912.66 ± 1095.269	486.428 (87.36, 5052.016)	0.022a
	Yes	344.597 ± 322.392	174.771 (57.341, 902.746)
Entropy	No	6.8 ± 0.474	6.759 (5.917, 7.557)	0.331
	Yes	6.527 ± 0.657	6.738 (5.409, 7.278)
Kurtosis	No	5.166 ± 2.863	4.547 (1.815, 11.855)	0.108
	Yes	3.367 ± 1.385	3.129 (2.072, 6.459)
25 percentile	No	0.276 ± 0.121	0.25 (0.156, 0.703)	0.933
	Yes	0.257 ± 0.195	0.253 (0, 0.674)
50 percentile	No	0.467 ± 0.166	0.417 (0.255, 0.817)	0.190
	Yes	0.524 ± 0.122	0.531 (0.332, 0.749)
75 percentile	No	0.67 ± 0.128	0.619 (0.433, 0.901)	0.035a
	Yes	0.77 ± 0.098	0.809 (0.596, 0.856)

FSOH: first and second-order histogram features; SD: standard deviation.

^aStatistically significant results.

Figure 4.

Examples of apparent diffusion coefficient (ADC) map histogram; x axis shows the brightness intensity and y axis is representative of frequency of each brightness intensity; histogram (a), depicts the distribution of low collagen contents of a macroadenoma whereas, histogram (b) shows a high-collagen contents macroadenoma. Distinct difference is obvious between these two histograms and this may be helpful in differential diagnosis of tumor's consistency and entities.

Table 3.

Comparison of apparent diffusion coefficient (ADC) values obtained from magnetic resonance (MR) imaging among three categorized groups according to the tumor's collagen contents to explore the ability of features for differentiating soft and solid types of tumors; obtained p-values, show that according to this type of categorization, any feature is not well for such purpose.

FSOH	Percentage of collagen	ADC mean±SD	ADC median (min, max)	p-Value
Maximum	<1%	2566.083 ± 330.477	2548.833 (2045, 3165)	0.886
	1–3%	2598.875 ± 567.985	2636.6 (1423, 3425)
	>3%	2514.843 ± 807.148	2783.9 (1353, 3138.571)
Mean	<1%	1187.437 ± 371.937	1121.684 (627.214, 1879.57)	0.747
	1–3%	1084.169 ± 300.413	1075.668 (691, 1864.688)
	>3%	1094.725 ± 567.148	1081.636 (423.46, 1792.17)
Median	<1%	1090.313 ± 399.121	1019.5 (557.5, 1876.5)	0.709
	1–3%	987.375 ± 343.13	912.5 (614.5, 1842)
	>3%	1053.375 ± 586.819	932.75 (485, 1863)
Minimum	<1%	498.554 ± 378.353	461.179 (76, 1557)	0.131
	1–3%	325.074 ± 136.425	361.375 (115, 512)
	>3%	191.936 ± 141.172	214.071 (0, 339.6)
Normal mean	<1%	0.417 ± 0.147	0.375 (0.237, 0.751)	0.673
	1–3%	0.364 ± 0.082	0.353 (0.26, 0.495)
	>3%	0.378 ± 0.15	0.306 (0.296, 0.603)
Standard deviation	<1%	0.156 ± 0.058	0.155 (0.07, 0.295)	0.4
	1–3%	0.173 ± 0.041	0.158 (0.135, 0.274)
	>3%	0.189 ± 0.05	0.18 (0.144, 0.253)
Smoothness	<1% 1–3% >3%	7.86 × 10−8±1.89 × 10−7 9.62 × 10−8±2.20 × 10−7 7.43 × 10−8±9.61 × 10−8	7.03 × 10−9 (5.59 × 10−12, 7.22 × 10−7) 3.43 × 10−9 (1.37 × 10−10, 6.26 × 10−7) 4.71 × 10−8 (3.00 × 10−10, 2.02 × 10−7)	0.837
Third Moment	<1%	0.004 ± 0.004	0.002 (–0.001, 0.019)	0.698
	1–3%	0.004 ± 0.004	0.005 (–0.003, 0.01)
	>3%	0.002 ± 0.004	0.003 (–0.003, 0.005)
Uniformity	<1%	888.739 ± 1278.754	372.989 (87.36, 5052.016)	0.864
	1–3%	576.615 ± 343.981	540.802 (57.341, 1034.986)
	>3%	495.105 ± 377.341	493.287 (91.099, 902.746)
Entropy	<1%	6.668 ± 0.584	6.845 (5.64, 7.325)	0.931
	1–3%	6.827 ± 0.368	6.756 (6.149, 7.557)
	>3%	6.625 ± 0.846	6.886 (5.409, 7.319)
Kurtosis	<1%	4.881 ± 3.116	3.523 (1.973, 11.855)	0.820
	1–3%	4.294 ± 1.892	3.773 (1.815, 7.709)
	>3%	3.995 ± 2.273	3.734 (2.054, 6.459)
25 Percentile	<1%	0.303 ± 0.176	0.289 (0, 0.703)	0.538
	1–3%	0.253 ± 0.068	0.25 (0.16, 0.351)
	>3%	0.208 ± 0.176	0.2 (0, 0.431)
50 Percentile	<1%	0.509 ± 0.158	0.499 (0.255, 0.817)	0.717
	1–3%	0.456 ± 0.131	0.43 (0.3, 0.698)
	>3%	0.491 ± 0.204	0.427 (0.332, 0.78)
75 Percentile	<1%	0.709 ± 0.138	0.742 (0.433, 0.901)	0.888
	1–3%	0.697 ± 0.115	0.693 (0.557, 0.856)
	>3%	0.7 ± 0.13	0.665 (0.591, 0.88)

SD: standard deviation; FSOH: first and second-order histogram.

Table 4.

Cut-off values, area under the curve (AUC), accuracy, sensitivity and specificity values for definitive cut-off points and p-values obtained from receiver-operator characteristic (ROC) analysis. As it is obvious from p-values, only cut-off points of smoothness, uniformity and 75 percentile were statistically significant.

FSOH	Cut off value	AUC	Accuracya	Sensitivity% (95% CI)	Specificity% (95% CI)	p-Value
Maximum	>2616	0.624	0.248	80 (44.4–97.5)	61.9 (38.4–81.9)	0.2688
Mean	≤691	0.538	0.076	20 (2.5–55.6)	95.24 (76.2–99.9)	0.7490
Median	≤831	0.581	0.162	50 (18.7–81.3)	76.19 (52.8–91.8)	0.4934
Minimum	≤193.16	0.619	0.238	50 (18.7–81.3)	90.48 (69.6–98.8)	0.3461
Normal Mean	>0.377	0.557	0.114	50 (18.7–81.3)	71.43 (47.8–88.7)	0.6142
Standard Deviation	>0.154	0.690	0.38	80 (44.4–97.5)	57.14 (34.0–78.2)	0.0890
Smoothness	>5.7 × 10−9	0.752	0.504	70 (34.8–93.3)	33.33 (14.6–57.0)	0.0134b
Third Moment	>0.0052	0.614	0.228	40 (12.2–73.8)	90.48 (69.6–98.8)	0.3474
Uniformity	≤179.271	0.757	0.514	60 (26.2–87.8)	90.48 (69.6–98.8)	0.0082b
Entropy	≤5.81787	0.610	0.22	30 (6.7–65.2)	100 (83.9–100)	0.3357
Kurtosis	<4.64	0.681	0.362	90 (55.5–99.7)	47.62 (25.7–70.2)	0.0635
25th Percentile	≤0.159	0.51	0.02	30 (6.7–65.2)	95.24 (76.2–99.9)	0.9421
50th Percentile	>0.444	0.648	0.296	80 (44.4–97.5)	66.67 (43.0–85.4)	0.1546
75th Percentile	>0.7	0.738	0.476	80 (44.4–97.5)	66.67 (43.0–85.4)	0.0088b

ADC: apparent diffusion coefficient; CI: confidence interval; FSOH: first and second order histogram features

^aAccuracy is calculated using (2×AUC)⁻¹ formula; ^bstatistically significant results.

Figure 5.

Receiver-operator characteristic (ROC) curves of first or second-order histogram (FSOH) features showing sensitivity and specificity of each variable. Uniformity, smoothness, and 75th percentile were significant FSOHs regarding ROC analysis (see Table 4) and for differential diagnosis of functional tumors from non-functional types.

Figure 6.

Additional receiver-operator characteristic (ROC) curves of first or second-order histogram (FSOH) features showing sensitivity and specificity of each variable (a) and (b). (c) represents a cumulative receiver-operator characteristic (ROC) curve for meaningful variables. As mentioned in the results, only smoothness, uniformity, 75th percentile were statistically significant for differentiation of functional from non-functional tumors.

Most features had lower ADC values in the group with collagen contents <3% than the group with collagen contents <1%. Anyway, statistical analysis did not show any significant differences for any FSOH feature represented in Table 3. Furthermore, Table 5, contains 2 × 2 crosstabs, diagnostic test PPV and NPV values. After categorizing tumors to the active (disease present) or inactive (disease absent) status, according to the cut-off values of smoothness, uniformity and 75th percentile, the PPVs and NPVs were calculated against tumor functionality. Smoothness has PPV = 30% (6.67% to 65.25%) and NPV = 85.71% (63.66% to 96.95%); uniformity has PPV = 60% (26.24% to 87.84%) and NPV = 90.48% (69.62% to 98.83%); and 75th percentile has PPV = 80% (44.39% to 97.48%) and NPV = 61.90% (38.44% to 81.89%) for the specified cut-off as mentioned in material and method section. As is obvious, smoothness and uniformity are more certifiable for negative results confirmation or diagnosis of non-functional tumors, whereas 75th percentile is more applicable for positive results or diagnosis of functional tumors.

Table 5.

The frequency of functional and non-functional tumors against active or inactive status; functionality is determined using histochemical staining and activity status is evaluated using cut-offs obtained from receiver-operator characteristic (ROC) analysis for smoothness, uniformity and 75th percentile. Using 2 × 2 tables, positive predictive value and negative predictive value were calculated. Cut-off values were smoothness >5.7058 × 10⁻⁹, uniformity ≤ 179.271 and 75th percentile > 0.7009; in these situations the disease is considered active. Note that the functionality of tumors were determined using pituitary-related hormones previously.

	Disease present (active cases)	Disease absent (inactive cases)	Positive predictive value% (95% CI)	Negative predictive value% (95% CI)
Functionality
Functional	3	7	30% (6.67% to 65.25%)	85.71% (63.66% to 96.95%)	Smoothness
Non-functional	3	18
Functional	6	4	60% (26.24% to 87.84%)	90.48% (69.62% to 98.83%)	Uniformity
Non-functional	2	19
Functional	8	2	80% (44.39% to 97.48%)	61.90% (38.44% to 81.89%)	75th Percentile
Non-functional	8	13

CI: confidence interval.

Discussion

DWI is a MRI-based method for monitoring tumor progression and response to treatment.²⁰ Diffusion-weighted images are evaluated using special software and modules to extract the statistical features from image textures. Recently, Razek²¹ has mentioned that “Advanced image analysis, including histogram analysis, texture analysis, and machine learning, has recently been introduced and still is in research”, (page 246).²¹ The textural analysis is done on the first/second-order statistics including mean, variance or SD, skewness, kurtosis, angular second moment, contrast, correlation, homogeneity, entropy and so forth.^21–24 In the current study, we have evaluated the diagnostic test accuracy of selected histogram-extracted features for differential diagnosis of functional from non-functional macro-adenoma. Firstly, we have shown that the FSOHs may accurately be applicable for differentiation of functional macro-adenoma tumors from non-functional types; secondly, we have proposed probable cut-off points of FSOH features which may be applicable in para-clinical decision when categorizing macro-adenoma tumors; we think that this may help neurosurgeons to select the type of operation; we have seen that FSOH features are not applicable for differentiation of high-content collagen macro-adenomas from low-content types.^8,25 Note that the collagen contents are determinative for the tumor cell characteristics. Investigators have emphasized the role of collagen types and contents in the metastatic process; even the orientation of collagen is reported to be important for invasive properties of tumors.^26,27

The functionality of pituitary adenomas is a multi-dimensional subject and a multifactorial property. This matter complicates the differential diagnosis of functional from non-functional tumors.^28,29 Functional macro-adenomas are associated with serious clinical manifestations in most cases, but non-functional types may have or not have this association.³⁰ Thus, differential diagnosis of both types is important in clinical evaluation or selection of therapeutic strategy.^4,5 DWI and low ADC values are suggested as the predictors of aggressive pituitary macro-adenoma.³¹ In the present study, we have seen that ADC values of histogram features including maximum, mean, median, uniformity, entropy, kurtosis and 25th percentile were lower in functional macro-adenomas compared to non-functional; however, from this feature, only uniformity had a significant result (Table 2; ADC mean±SD values). Further, smoothness was meaningfully lower in the functional group than non-functional (Table 2). Researchers have reported discrepant results regarding the usefulness of MRI for predicting tumor consistency. For example, Alimohamadi et. al.,⁸ Yrjänä and coworkers,³² have reported a good correlation between tissue hardness with image intensity at the lesion sites. On other hand, Chakraborty and colleagues,⁹ Bahuleyan et. al.³³ and Suzuki and coworkers³⁴ have not found any correlation between signal intensity or ADC values with tumor consistency; their findings confirm our results. Such diverse findings may be due to the entity, physiological, and functionality or histologic status of the tumors in different studies. Further, the MRI machine and its capabilities for imaging and image analysis software installed on the machine and optimization of the imaging technique may be important. Wei and coworkers have stated that DWI is susceptible to the status of tissue consistency and errors arising from patient movement during imaging process.³⁵ Xing and coworkers have justified the parameters for diagnosis of pituitary macro-adenomas at 1.5 T magnetic resonance.³⁶ Some evidence has shown that the collagen contents are reversely correlated with ADC values; however, this research emphasizes that more studies should be done for more precise results.¹² In addition, in our evaluation, most FSOH features had lower ADC values in the group with collagen contents >3% than the group with collagen contents <1%. Anyway, statistical analysis did not show any significant differences for any FSOH feature represented in Table 3. We think that informative data of ROC analysis, which are presented in the current study, should be validated in the other centers across the world because of their scarcity. In fact, we did not find any investigation with the considerations we offer in this research. In the present study, the best sensitivity and specificity cut-off values for each FSOH feature are reported. Image analysis methods, are of interest in the field of neuroradiology; however, nowadays labeling methods are of interest to investigators. For example, the combination of arterial spin labeling (ASL) perfusion and diffusion tensor imaging (DTI)-derived metrics is reported to be a suitable noninvasive method for differential diagnosis of residual or recurrent gliomas from post-radiation changes.³⁷

In the present study, we have used selected FSOHs as the extracted features from histograms which are more tangible for clinicians and more straightforward to calculate; furthermore, specifying cut-off values on histogram features is reasonable for probable decision-making purposes. But, for the establishment of selected features, more studies with higher sample sizes may be needed.

Insignificant results of comparison of histogram-extracted features among three categorized groups according to the collagen contents, implies the inability of using FSOH for tumor consistency differential diagnosis (see Table 3). However, it may be possible that a meaningful result could be obtained using a higher sample size; however, there were only four patients with >3% collagen contents using the Masson trichromatic staining method, which may affect the results of the study. Furthermore, more advanced technologies for DWI in future years may potentiate the diagnostic values of FSOH features. Another cause of insignificancy of mentioned results could be the abnormal distribution of FSOH features and, again, higher sample sizes may resolve this flaw. Razek et. al. have evaluated the ADC values obtained from soft tissue tumors of the extremities using diffusion echo-planar MRI on 37 patients; they have reported that ADC mean±SD in malignant and benign tumors mass were 1.02 ± 0.03 × 10⁻³ mm²/s and 1.54 ± 0.03 × 10⁻³ mm²/s, respectively (p < 0.001). They have selected 1.34 × 10⁻³ mm²/s as the cut-off value for differential diagnosis of malignant tumors from benign types and reported an the accuracy of 91%, sensitivity of 94% and 94% specificity. They have discussed that diffusion-weighted-echo-planar MRI is a helpful non-invasive modality for differentiation of malignant soft tissue tumors from benign types.²⁴ Compared to Razek and coworkers, we evaluated more detailed features and have shown that only three ADC histogram-extracted features are applicable for differential diagnosis of functional tumors from non-functional types. Misdiagnosis in differentiation of malignant and benign tumors is reported when using ADC value due to the arrangement of collagen fibers surrounding the tumor cells.²⁴

All cut-off values presented in Table 4, include AUCs >50% but the most impressive data belong to the smoothness, uniformity and 75th percentile. A practical discussion could be that: considering three variables with specified cut-offs, concurrently, the smoothness >5.7 × 10⁻⁹, the uniformity ≤179.271 and the 75th percentile >0.7, for a taken diffusion-weighted image that is processed for FSOH, a neuroradiology specialist may propose helpful comments about the tumor functionality to a neurosurgeon who should make surgical interventions.

Using obtained cut-offs for significant FSOH features and based on PPV and NPV values presented in Table 5, we suggest that the application of smoothness and uniformity for diagnosis purposes and in clinical settings are certifiable for negative results confirmation or diagnosis of non-functional tumors. Whereas, 75th percentile is appreciated for positive results or diagnosis of functional tumors.

In conclusion, DWI or any type of enhancement method followed by histogram feature analysis could result in a valuable differential diagnosis index which may play an important role in prognosis prediction or clinical decision for selection of an intervention method. Current work showed that DWI and histogram-extracted smoothness, uniformity and 75th percentile are potential differentiating indices for categorizing functional macro-adenomas from non-functional types. In addition, in the present study, we did not see any association between selected histogram-extracted features and collagen contents.

Informed consent

Informed consent was obtained from all individual participants included in the study.

Funding

This study was funded by Shahid Beheshti University of Medical Sciences, Tehran, Iran.

Conflict of interest

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Ethical approval

All procedures performed in the studies involving human participants were in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki Declaration and its later amendments or comparable ethical standards.