Quality reporting of radiomics analysis in pituitary adenomas: promoting clinical translation

Abstract

Objective:

To evaluate the quality of radiomics studies on pituitary adenoma according to the radiomics quality score (RQS) and Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD).

Methods:

PubMed MEDLINE and EMBASE were searched to identify radiomics studies on pituitary adenomas. From 138 articles, 20 relevant original research articles were included. Studies were scored based on RQS and TRIPOD guidelines.

Results:

Most included studies did not perform pre-processing; isovoxel resampling, signal intensity normalization, and N4 bias field correction were performed in only five (25%), eight (40%), and four (20%) studies, respectively. Only two (10%) studies performed external validation. The mean RQS and basic adherence rate were 2.8 (7.6%) and 26.6%, respectively. There was a low adherence rate for conducting comparison to “gold-standard” (20%), multiple segmentation (25%), and stating potential clinical utility (25%). No study stated the biological correlation, conducted a test–retest or phantom study, was a prospective study, conducted cost-effectiveness analysis, or provided open-source code and data, which resulted in low-level evidence. The overall adherence rate for TRIPOD was 54.6%, and it was low for reporting the title (5%), abstract (0%), explaining the sample size (10%), and suggesting a full prediction model (5%).

Conclusion:

The radiomics reporting quality for pituitary adenoma is insufficient. Pre-processing is required for feature reproducibility and external validation is necessary. Feature reproducibility, clinical utility demonstration, higher evidence levels, and open science are required. Titles, abstracts, and full prediction model suggestions should be improved for transparent reporting.

Advances in knowledge:

Despite the rapidly increasing number of radiomics researches on pituitary adenoma, the quality of science in these researches is unknown. Our study indicates that the overall quality needs to be significantly improved in radiomics studies on pituitary adenoma, and since the concept of RQS and IBSI is still unfamiliar to clinicians and radiologist researchers, our study may help to reach higher technical and clinical impact in the future study.

Introduction

Pituitary adenomas are the third most common primary brain tumors and comprise 16.7% of all brain tumors. ¹ Although pituitary adenomas are rarely malignant, they may significantly disrupt the quality of life and often require medical interventions, surgery, radiation treatment, and life-long follow-up. ² MRI remains the gold-standard of imaging for diagnosis and follow-up combined with biochemical tests and visual testing. ^3–6 Tumor volume and the location relationship with adjacent structures (i.e. optic chiasm, cavernous sinus, etc.) need to be evaluated with MRI. MRI also plays an important role in assessing the therapeutic response of pituitary adenomas by tracking tumor volume changes. ⁷

Radiomics is an emerging field for extracting quantitative, mineable, and high-dimensional data from medical images to support decision-making. The underlying concept is that quantitative features (i.e. shape, first-order, and second-order [texture] features) may reflect the underlying pathophysiology of the tissue. ⁸ Previous radiomics studies on neuroradiology have mostly focused on glioma. ^9–11 However, an increasing number of recent radiomics studies have shown promising findings on diverse tasks in pituitary adenomas; specifically, differential diagnosis, ^12–21 as well as prediction of underlying pathology, ^22–25 response to treatment, ^26–29 and recurrence or progression. ^30,31

For radiomics to overcome the current translational gap between an exploratory research and a clinically approved tool for decision-making, there is a need for methodology standardization for reproducibility and evaluation of the clinical-biomarker and biomarker-outcome correlations. ³² Recently, a radiomics quality score (RQS) was proposed as useful for assessing the quality of studies. ³³ There have been previous studies on RQS in oncology ^34,35 and glioma. ³⁶ Furthermore, the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD) initiative is also desirable, because radiomics research is a model-based approach. To the best of our knowledge, the quality of science in radiomics studies on pituitary adenoma remains unclear.

Therefore, this study aimed to evaluate the quality of radiomics studies on pituitary adenoma using RQS and TRIPOD. Specifically, we aim to discuss the current challenges underlying the translational gap in radiomics for imaging biomarkers of pituitary adenomas, as well as to postulate for future clinical application.

Methods and materials

Systematic search strategy and study selection

PubMed MEDLINE (n = 64) and EMBASE (n = 74) databases were used to search and collect all original research papers using radiomics analysis, published until September 17, 2020 and the following terms were used for search: (“pituitary adenoma” OR “pituitary macroadenoma” OR “prolactinoma” OR “acromegaly” OR “Cushing”) AND (“radiomic” OR “texture”). Of the total 138 papers searched, 38 duplicated articles were removed, and 79 were further excluded for the following reasons: non-radiomics studies (n = 34), not in field of interest (n = 29), non-brain image (n = 9), conference abstracts (n = 5), review article (n = 1), and editorial (n = 1). Of the remaining 21 articles, a study with main text in language other than English (n = 1) was excluded. Finally, 20 articles were used for analysis (Figure 1 and Supplementary Material 1).

Figure 1.

Flow chart of the study selection process.

Analysis of method quality based on RQS

The RQS score consisted of a total 16 components and was classified into 6 domains to evaluate RQS, as reported in previous studies (Supplementary Material 1). ^35,36 Before RQS evaluation, the reviewers conducted a meeting to achieve consensus on the evaluation criteria. Two reviewers (with 6 years and 9 years of experience in radiology, respectively) independently evaluated RQS for the articles for each of the six domains (Supplementary Material 1). In case of disagreement, a senior radiologist (with 15 years of experience) was consulted for the final decision.

For additional RQS scoring, which require further discussion due to nature of the studies on pituitary adenoma, a consensus was reached through discussion of the reviewers. For RQS, additional discussion and consensus were made for the following topics: issues for ‘validation’ (domain 2), ‘multivariate analysis with non-radiomics features’ (domain 3), and ‘comparison with gold-standard’ (domain 3) (Supplementary Material 1).

Analysis of reporting completeness based on TRIPOD statement

The TRIPOD checklist, consisting of 37 items in 22 main criteria, was applied to each article to determine the completeness of reporting. ³⁷ The details for TRIPOD checklist and data extraction are shown in Supplementary Material 1.

Statistical analysis

The characteristics of total 20 articles were reviewed. In case a score of even one point for each item (0–16 items) is obtained, it was considered to have basic adherence for each item to RQS. The basic adherence was calculated, and the basic adherence rate (%) was calculated as proportion of the number of articles with basic adherence to number of total articles. The RQS score for each item was reported as means and standard deviation. Percentage of the ideal score (%) was calculated as proportion of mean score to ideal score for each item, and the total RQS score (-8 to 36) was counted for all articles. Total 35 items in TRIPOD was scored and adherence to individual TRIPOD items was described using proportion (%). When the overall adherence rate was calculated, validation items (10c, 10e, 12, 13, 17 and 19a) and “if done” item (5c, 11) were excluded from both the denominator and numerator.

Results

Characteristics of the 20 included radiomics studies in pituitary adenomas

Table 1 and Figure 2 present the characteristics of the 20 included radiomics studies. ^12–31 The median number of patients included in the studies was 84 (range: 27–235). Regarding the journal type, there were 7 (35%), 11 (55%), and 2 (10%) clinical, ^{12,21,22,24,26,27,31} imaging, ^{13–15,17–20,23,25,28,29} and computer/neuroscience journals, ^16,30 respectively. Radiomics studies were diagnostic (70%), ^12–25 predictive (20%), ^26–29 or prognostic (10%). ^30,31 The study topics included differential diagnosis (50%), ^12–21 recurrence/progression prediction (10%), ^30,31 pathologic classification (20%), ^22–25 and response to treatment (20%). ^26–29 The assessed tumor subtypes included the entire pituitary adenoma (50%), ^13–21,25 acromegaly (30%), ^{12,22,24–27,29} non-functioning pituitary adenoma, (15%), ^23,30,31 and invasive functional pituitary adenoma (5%). ²⁸

Table 1.

Characteristics of the 20 included radiomics studies

Article characteristics	No. of articles
No. of subjects	84 (27–235)
Journal type
Clinical journal	7 (35)
Imaging journal	11 (55)
Computer science/Neuroscience journal	2 (10)
Biomarker
Diagnostic	14 (70)
Predictive	4 (20)
Prognostic	2 (10)
Topics in pituitary adenoma
Differential diagnosis	10 (50)
Prediction of recurrence/progression	2 (10)
Pathologic classification	4 (20)
Response to treatment	4 (20)
Assessed tumor subtype
Entire pituitary adenoma	10 (50)
Acromegaly	6 (30)
NFPA	3 (15)
IFPA	1 (5)
Preprocessing performed
Isovoxel resampling	5 (25)
N4 bias field correction	4 (20)
Signal intensity normalization	8 (40)
Sequence used for feature extraction
Conventional images	18 (90)
Advanced images (Perfusion or diffusion)	2 (10)
Region of interest in tumor segmentation
Placement of small regions of interest	1 (5)
Tumor on a single 2D slice	5 (25)
Whole tumor on entire slice	11 (55)
N/A	3 (15)
Software for feature extraction
Pyradiomics	10 (50)
LIFEx	2 (10)
Matlab	3 (15)
ImageJ	2 (10)
Omni-kinetics	2 (10)
N/A	1 (5)
Segmentation method
Manual	18 (90)
Semi-automatic	2 (10)
External validation
Performed	2 (10)
Not performed	18 (90)
Magnetic field strength (Tesla)
1.5	2 (10)
3.0	14 (70)
1.5 & 3.0	4 (20)

IFPA, Invasive functional pituitary adenoma; N/A, not available; NFPA, non-functioning pituitary adenoma.

Numbers in parentheses are either interquartile range or percentages.

Figure 2.

Summary chart of the radiomics studies, according to the (a) topics in pituitary adenomas, (b) assessed tumor subtypes, (c) segmentation method, (d) region of interest in tumor segmentation. IFPA, invasive functional pituitary adenoma; NFPA, non-functioning pituitary adenoma.

Most of the studies did not perform preprocessing; specifically, only 5 (25%), ^{13,14,17,22,29} 4 (20%), ^17,22,29,30 and 8 (40%) ^{13,14,17,22,23,25,29,30} studies performed isovoxel resampling, N4 bias field correction, and signal intensity normalization, respectively. Most studies (90%) used conventional images (non-contrast T ₁ weighted, T ₂ weighted, or contrast-enhanced T ₁ weighted images) for feature extraction while the remaining studies (10%) ^19,20 used advanced images, including perfusion, or diffusion images for feature extraction. To delineate the region of interest in tumor segmentation, 1 (5%) ¹⁷ study placed small regions of interest instead of including the whole tumor, 5 (25%) ^{13,14,18,26,31} studies only segmented a part of the tumor on a single two-dimensional slice, and 11 (55%) ^{15,16,19,21,22,24,25,27–30} studies segmented the whole three-dimensional tumor volume on entire slices. The other three studies (15%) ^12,20,23 did not provide information about ROI. The software used for radiomics feature extraction included Pyradiomics (50%), ^{12–14,17,22,24,25,27,29,30} LIFEx (10%), ^16,21 Matlab (15%), ^15,23,28 Image J (10%), ^26,31 and Omni-kinetics (10%). ^19,20 Among them, Pyradiomics and LIFEx adhere to the Image Biomarker Standardization Initiative (IBSI). ³⁸ Regarding segmentation, manual segmentation (90%) ^{12–21,23–29,31} was mainly used, followed by semi-automatic segmentation (10%). ^22,30 Only two (10%) ^12,24 studies performed external validation. 14 (70%), 4 (20%), and 2 (10%) studies used 3.0 Tesla magnets, 1.5 Tesla magnets, and both, respectively, for radiomics analysis.

Basic adherence rate of the reporting quality according to the six key domains

Table 2 summarizes the basic adherence rates to the RQS for 16 items. Regarding domain 1, 16 (80%) studies used well-documented image protocols. ^{12–15,17–23,25,27–30} None of the studies performed imaging scans at multiple time points to acquire reliability and phantom study. Two (10%) studies performed semi-automatic segmentation. ^22,30

Table 2.

Radiomics quality score according to the six key domains

	Basic adherence rate (%)	Mean score (mean ± standard deviation)	Percentage of the ideal score (%)
Total 16 items (ideal score 36)	85 (26.6)	2.8 ± 8.7	7.6
Domain 1 – protocol quality and stability in image and segmentation (0–5 points)	21 (26.3)	1.1 ± 0.7	21.0
Protocol quality (2)	16 (80.0)	0.8 ± 0.4	40.0
Test-retest (1)	0 (0)	0	0
Phantom study (1)	0 (0)	0	0
Multiple segmentation (1)	5 (25.0)	0.3 ± 0.4	25.0
Domain 2 – feature selection and validation (-8 to 8 points)	20 (50.0)	−1.3 ± 6.0	−16.3
Feature reduction or adjustment of multiple testing (−3 or 3)	12 (60.0)	0.6 ± 3.0	20.0
Validation (−5, 2, 3, 4, or 5)	8 (40.0)	−1.9 ± 4.0	−38.0
Domain 3 – biologic/clinical validation and utility (0–6 points)	15 (18.8)	1.2 ± 2.1	20.0
Non-radiomics features (1)	6 (30.0)	0.3 ± 0.5	30.0
Biologic correlations (1)	0 (0)	0.0 ± 0.0	0.0
Comparison to “gold-standard” (2)	4 (20.0)	0.4 ± 0.8	20.0
Potential clinical utility (2)	5 (25.0)	0.5 ± 0.9	25.0
Domain 4 – model performance index (0– 5 points)	29 (48.3)	1.8 ± 1.3	36.0
Cut-off analysis (1)	10 (50.0)	0.5 ± 0.5	50.0
Discrimination statistics (2)	13 (65.0)	0.9 ± 0.8	45.0
Calibration statistics (2)	6 (30.0)	0.4 ± 0.7	20.0
Domain 5 – high level of evidence (0–8 points)	0 (0)	0	0
Prospective study (7)	0 (0)	0	0
Cost-effectiveness analysis (1)	0 (0)	0	0
Domain 6 – open science and data (0–4 points)	0 (0)	0	0

Data are the number of studies. Numbers in parentheses are percentage.

Regarding domain 2, 12 (60%) studies performed feature reduction or adjustment for multiple testing. ^{12–15,17,23–25,27,28} Six (30%) studies performed validation based on a data set obtained from the same institution and gained +2 points. ^{13–16,23,28} Only two (10%) studies performed validation based on three or more data sets from distinct institutions and gained a full score of +5. ^12,24 The remaining 12 studies (60%) did not perform any form of validation.

In domain 3, six (30%) studies performed multivariate analysis with non-radiomics features to develop a differential diagnosis model, pathologic classification model, and treatment response prediction model. ^{12,15,23,24,27,28} No study evaluated the biological correlation. Four (20%) studies performed a comparative analysis with “gold-standard”, including Ki-67 or hormonal level. ^12,24,27,28 Five (25%) studies confirmed clinical current and potential utility through decision curve analysis for diagnosis of tumor consistency and cavernous sinus invasion, prediction of treatment response, and prediction of Ki-67; thus, they received two points in the “clinical utility” component. ^{12,15,24,27,28}

In domain 4, 10 (50%) studies performed cut-off analysis using a reference standard cut-off value. ^{13,15,17,19,24,26–28,31} 13 (65%) studies performed a resampling method (including cross-validation or bootstrapping) or discrimination statistics (including receiver operating characteristic curve or area under the curve) to determine their statistical significance. ^{12,14,15,17,18,20–25,27,28} Six (30%) studies reported calibration statistics with their significance or resampling methods, including cross-validation or bootstrapping. ^{12,15,23,24,27,28}

In domain 5, none of the studies were prospective or performed cost-effective analysis. Finally, in domain 6, none of the studies were publicly available.

RQS assessment

The mean overall RQS score for the 20 radiomics studies was 2.8 ± 8.7, which was 7.6% of the ideal score (Table 2 and Figure 3). The highest and lowest scores were 18 and −7, respectively, which corresponded to 50% and −19.4% of the ideal score, respectively. Among the six domains, domain 2 (feature selection and validation) and domain 4 (model performance) had the lowest (−1.3 and −16.3%) and highest (1.8 and 36%) mean scores and percentages of the ideal score, respectively.

Figure 3.

RQS assessment results according to the six key domains. RQS, radiomics quality score.

Seven (35%) studies ^{18–21,26,30,31} did not perform feature selection or validation; among them, three studies had the lowest RQS. ^26,30,31 The study with the highest score obtained the ideal score for protocol quality, feature reduction or adjustment for multiple testing, validation, non-radiomics features, comparison to the “gold-standard”, potential clinical utility, discrimination statistics, and calibration items.

Completeness in reporting a radiomics-based multivariable prediction model using TRIPOD assessment

Upon the inclusion of all 35 items, the mean number of reported TRIPOD items was 15.7  ±  3.85 (standard deviation; range: 11–24). Upon exclusion of “if relevant” and “if done” items from both the numerator and denominator, the adherence rate for TRIPOD was 54.6%. Table 3 demonstrates the adherence rates to individual TRIPOD items.

Table 3.

Adherence to individual TRIPOD items in radiomics study

Total (35 item)	All article (n = 20)
Title and abstract
Title
1. Identify developing/validating a model, target population, and the outcome	1 (5)
Abstract
2. Provide a summary of objectives, study design, setting, participants, sample size, predictors, outcome, statistical analysis, results, and conclusions	0 (0)
Introduction
Background and objectives
3a. Explain the medical context and rationale for developing/validating the model	20 (100)
3b. Specify the objectives, including whether the study describes the development/validation of the model or both	6 (30)
Methods
Source of data
4a. Describe the study design or source of data (randomized trial, cohort, or registry data)	3 (15)
4b. Specify the key dates	19 (95)
Participants
5a. Specify key elements of the study setting including number and location of centers	13 (65)
5b. Describe eligibility criteria for participants (inclusion and exclusion criteria)	20 (100)
5c. Give details of treatment received, if relevant (n  =  2)	two out of 2
Outcome
6a. Clearly define the outcome, including how and when assessed	16 (80)
6b. Report any actions to blind assessment of the outcome	5 (25)
Predictors
7a. Clearly define all predictors, including how and when assessed	19 (95)
7b. Report any actions to blind assessment of predictors for the outcome and other predictors	9 (45)
Sample size
8. Explain how the study size was arrived at	2 (10)
Missing data
9. Describe how missing data were handled with details of any imputation method	5 (25)
Statistical analysis methods
10a. Describe how predictors were handled	20 (100)
10b. Specify type of model, all model-building procedures (any predictor selection), and method for internal validation	14 (70)
10d. Specify all measures used to assess model performance and if relevant, to compare multiple models (discrimination and calibration)	6 (30)
Risk groups
11. Provide details on how risk groups were created, if done (yes or no, n  =  0)	N/A
Results
Participants
13a. Describe the flow of participants, including the number of participants with and without the outcome. A diagram may be helpful	19 (95)
13b. Describe the characteristics of the participants, including the number of participants with missing data for predictors and outcome	15 (75)
Model development
14a. Specify the number of participants and outcome events in each analysis	5 (25)
14b. Report the unadjusted association between each candidate predictor and outcome, if done (yes or no, n =  20)	3 (15)
Model specification
15a. Present the full prediction model to allow predictions for individuals (regression coefficients, intercept)	1 (5)
15b. Explain how to the use the prediction model (nomogram, calculator, etc)	7 (35)
Model performance
16. Report performance measures (with confidence intervals) for the prediction model	10 (50)
Discussion
Limitations
18. Discuss any limitations of the study	20 (100)
Interpretation
19b. Give an overall interpretation of the results	19 (95)
Implications
20. Discuss the potential clinical use of the model and implications for future research	18 (90)
For validation (Types 2a, 2b, 3, and 4)	n = 8
Methods—Statistical analysis methods
10c. Methods—Statistical analysis methods: describe how the predictions were calculated	4 (50)
10e. Describe any model updating (recalibration), if done (n = 0)	0 (0)
Methods
12. Identify any differences from the development data in setting, eligibility criteria, outcome, and predictors	6 (75)
Results
13c. Show a comparison with the development data of the distribution of important variables	5 (62.5)
Results—model updating
17. Report the results from any model updating, if done (n = 0)	0 (0)
Discussion—interpretation
19a. Discuss the results with reference to performance in the development data and any other validation data	3 (37.5)

N/A, not available.

Numbers in parentheses are percentages.

Discussion

Given the rapid increase in radiomics studies on pituitary adenoma, there is a need for a comprehensive evaluation of the research quality to promote clinical translation. Radiomics studies on pituitary adenomas were suboptimal with respect to the science and reporting quality, with an average score of 26.6 and 54.6% in the RQS and TRIPOD guidelines, respectively. Our findings indicate that radiomics studies on pituitary adenomas had insufficient methodological quality, requiring significant improvement to become a clinically applicable tool.

Radiomics research on pituitary adenomas has distinctive features from those of radiomics research on other brain diseases. ^39,40 First, compared with other brain diseases with open source data sets, including The Cancer Imaging Archive (TCIA) for glioma or Alzheimer’s Disease Neuroimaging Initiative (ADNI) for Alzheimer’s disease, ⁴¹ pituitary adenomas lack a well-labeled open source public dataset. This resulted in low scores of validation (−1.9) for domain 2 and of open science and data (0) for domain 6; moreover, most of the included studies were single-centered. To improve radiomics research on pituitary adenomas, there is a need to establish an open source data set for testing and possible clinical translation. Second, the small size of pituitary adenomas may limit radiomics application. Information obtained from second-order features may be limited in the small region size; consequently, several radiomics studies on pituitary adenomas have focused on pituitary macroadenomas (>1 cm) ^{12,14,17,19–21,23,24,26,30} rather than including the entire micro- and macroadenomas. Third, pituitary adenoma segmentation was more often manual than semi-automatic; moreover, different reviewers did not perform segmentation, which limits the reproducibility and generalizability of radiomics features. The predominance of manual segmentation could result from the unique pituitary gland anatomy. Specifically, since the pituitary gland is located outside the blood–brain barrier, ⁴² there is contrast enhancement around the normal pituitary gland. The lack of contrast-enhanced boundaries impedes semi-automatic delineation of the delayed enhancement of the pituitary adenoma. Contrastingly, other brain tumors, including gliomas or meningiomas, ⁴³ present a relatively distinctive contour compared to the background parenchyma. A practical approach may involve independent tumor segmentation by multiple readers and the performance of interobserver agreement analysis. Currently, only 5 (25%) studies have performed multiple segmentation.

We applied the previously designed six key domains, which supports RQS integration. ³⁵ Most of the key domains require significant improvement. Regarding the technical validation in domain 1, no studies performed test–retest or phantom studies, which indicates the overall insufficiency of data regarding the precision or technical bias. Additionally, the basic adherence rates of domain 2 (feature selection and validation; 50.0%) and domain 3 (biological/clinical validation and utility; 18.8%) were substantially lower than those reported by studies on glioma and oncology (81.4 and 39.2%, respectively). ³⁵ In domain 2, the basic adherence rate for feature selection was 60%, which was substantially lower than the previously reported rates by studies on glioma and oncology (94.1 and 96.1%, respectively). ³⁵ Radiomics studies lacking appropriate feature reduction are susceptible to overfitting. ⁴⁴ Similarly, the basic adherence rate of the validation in domain 2 was 40%, which was lower than the rates reported by previous glioma and oncology studies (68.6 and 70.1%, respectively), ³⁵ with the mean RQS score being below zero (−1.9). 12 (60%) studies lacked any validation type (60%). Moreover, only two (10%) studies performed external validation. The lack of proper validation may have likely produced overoptimistic results on performance. Regarding domain 3, no study correlated radiomics features with biological features. For domain 5, none of the studies were prospective or analyzed cost-effectiveness. On the other hand, for domain 6, none of the studies provided data, code, or segmentation mask in open source. Multicenter studies are required to allow reproducibility of the radiomics technique. Furthermore, releasing data and code in open source can accelerate radiomics field development.

The adherence rate to TRIPOD guidelines, which generally covers prediction models, was 54.4%, which was similar to that reported by oncology radiomics studies. ³⁵ Complete reporting of the prediction model studies allows study replication and study impact evaluation, which could facilitate the clinical application of prediction models. Inadequate reporting impedes the use of all available evidence regarding a prediction model. The title and abstract were among the least well-reported items. Several items should be present in the abstract; among them, one was missing in all studies. Specifically, they did not clearly describe “development” or “validation” in the study objective. Moreover, there were poor explanations regarding the determination of the sample size and suggestions of a full prediction model. Our results suggest the need to improve the transparent reporting of radiomics models.

Imaging pre-processing is a particularly crucial step for standardizing heterogeneous imaging data before feature extraction since radiomics feature preprocessing increases reproducibility. ^44,45 Lack of reproducible features increases the risk to Type I error or false discoveries. ⁴⁶ Isovoxel resampling is comprised in heterogeneous spatial resolution to maintain rotational invariability; contrastingly, bias field correction is frequently applied in MRI to mitigate distortions or non-uniformities that drastically alter the extracted features and hamper reproducibility. ⁴⁷ Signal intensity normalization is crucial in MRI radiomics analysis for reducing variance in the arbitrary signal intensity of conventional MRI sequences. A previous study demonstrated that lacking image intensity normalization causes classification errors, especially in second-order features. ⁴⁸ However, most radiomics studies on pituitary adenomas lacked descriptions regarding preprocessing steps in isovoxel resampling, bias field correction, and signal intensity normalization (80%, 60%, and 75%, respectively). Future studies should abide by the pre-processing steps for greater reproducibility.

The IBSI recently attempted to address the lack of standardization in radiomics by defining standardized nomenclature and producing tools for verifying radiomics software implementations, ³⁸ which resulted in a validated set of consensus-based reference values for radiomics features. Only 12 (60%) radiomics studies on pituitary adenomas provided descriptions regarding adherence of radiomics software application to the IBSI recommendation. To allow the clinical application of radiomics models, future studies should adopt a standardized approach complying with the IBSI standard.

This study has several limitations. First, there were relatively few pituitary adenoma articles on radiomics analysis. Second, as RQS is an expert opinion, some items are too idealistic; e.g. in real-world clinical settings, phantom study and multiple imaging acquisitions are rarely available. Nonetheless, future clinical translation of radiomics researches requires a higher reporting quality.

Conclusion

There is insufficient reporting quality in radiomics studies on pituitary adenomas. Preprocessing is required for feature reproducibility; additionally, external validation is necessary. There is a need to improve feature reproducibility, demonstration of clinical utility, higher evidence level, and open science. Regarding transparent reporting, the title, abstract, and suggestion of a full prediction model should be improved.