Professional Documents
Culture Documents
Peritumoural MRI Radiomics Signature of Brain Meta
Peritumoural MRI Radiomics Signature of Brain Meta
Clinical Radiology
journal homepage: www.clinicalradiologyonline.net
art icl e i nformat ion AIM: To investigate whether magnetic resonance imaging (MRI) radiomics features of brain
metastases (BMs) can predict epidermal growth factor receptor (EGFR) mutation status in lung
Article history: adenocarcinoma.
Received 4 May 2023 MATERIALS AND METHODS: Between June 2014 and December 2022, 58 histopathologically
Received in revised form confirmed lung adenocarcinoma patients (27 with EGFR wild-type, 31 with EGFR mutation)
5 October 2023 who underwent gadobenate dimeglumine-enhanced brain MRI were recruited retrospectively.
Accepted 18 October 2023 A total of 123 metastatic brain lesions were allocated randomly into the training cohort (n¼86)
and test cohort (n¼37) at a ratio of 7:3. Radiomics models based on multi-sequence MRI im-
ages in different regions such as volume of interest (VOI)enhancing tumour, VOIwholetumour,
VOIperitumour 1mm, VOIperitumour 3mm, and VOIperitumour 5mm were built. The optimal radiomics
model was integrated into the clinical or radiological indicators to construct a fusion model
through multivariable logistic regression analysis.
RESULTS: The optimal radiomics model based on the VOIperitumour 1mm, a combination of nine
features selected from the fluid-attenuated inversion recovery (FLAIR) sequence, yielded areas
under the curves (AUCs) of >0.75 in the training and test cohorts. The prediction of the fusion
model with integration of clinical factors (age) and radiomics score (the optimal radiomics
model) was not better than that of the optimal radiomics model alone in the test cohort (AUC:
0.808 and 0.785, respectively, p¼0.525).
CONCLUSION: The FLAIR radiomics model based on VOIperitumour 1mm as an effective
biomarker helps predict EGFR mutation status in lung adenocarcinoma patients with BMs and
then assists clinicians in selecting optimal treatment strategies.
Ó 2023 The Authors. Published by Elsevier Ltd on behalf of The Royal College of Radiologists.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/
licenses/by-nc-nd/4.0/).
* Guarantor and correspondent: Z. Huang, Department of Radiology, Longyan First Affiliated Hospital of Fujian Medical University, No. 105 North 91 Road,
Xinluo District, Fujian, 364000, China. Tel.: þ8613850600215.
E-mail address: tuxuezhao@163.com (Z. Huang).
https://doi.org/10.1016/j.crad.2023.10.022
0009-9260/Ó 2023 The Authors. Published by Elsevier Ltd on behalf of The Royal College of Radiologists. This is an open access article under the CC BY-NC-ND license (http://
creativecommons.org/licenses/by-nc-nd/4.0/).
e306 Z. Huang et al. / Clinical Radiology 79 (2024) e305ee316
Table 1
Magnetic resonance imaging (MRI) protocol.
Repetition time (ms) Echo time (ms) Matrix Voxel size (mm)
GE Signa HDxt 13 1.5 T1WI 5,000 107 320 192 0.75 0.75 6.00
T2WI 1,800 24 320 192 0.75 0.75 6.00
FLAIR 8,600 120 288 192 0.75 0.75 6.00
DWI 8,600 120 289 192 0.75 0.75 6.00
CE-T1WI 1,750 24 320 256 0.75 0.75 6.00
Philips Prodiva 3 1.5 T1WI 585 13 252 179 0.78 1.10 6.00
T2WI 2,500 120 368 368 0.60 0.60 6.00
FLAIR 7,500 100 244 175 0.90 1.13 6.00
DWI 3,000 82 128 110 1.72 2.00 6.00
CE-T1WI 161 2.3 356 273 0.65 0.75 6.00
Philips Achieva 17 3.0 T1WI 2,000 20 288 216 0.76 0.92 7.00
T2WI 1,991 90 400 240 0.57 0.83 7.00
FLAIR 11,000 125 352 197 0.65 0.93 7.00
DWI 2,877 50 192 90 1.20 2.56 7.00
CE-T1WI 260 4.6 400 255 0.57 0.72 5.00
Philips Ingenia II 20 3.0 T1WI 1,900 20 308 173 0.75 1.07 7.00
T2WI 4,000 104 356 356 0.65 0.65 7.00
FLAIR 9,000 101 260 185 0.70 1.23 7.00
DWI 2,314 90 144 110 1.60 2.09 7.00
CE-T1WI 257 4.6 400 255 0.57 0.72 5.00
Philips Ingenia cx 5 3.0 T1WI 2,000 20 356 201 0.65 0.94 5.00
T2WI 6,278 95 272 272 0.85 0.85 5.00
FLAIR 8,000 125 352 173 0.65 1.05 5.00
DWI 3,020 86 152 98 1.51 2.35 5.00
CE-T1WI 272 2.3 356 164 0.90 1.12 5.00
T1WI, T1-weighted imaging; T2WI, T2-weighted imaging; FLAIR, fluid-attenuated inversion recovery; DWI, diffusion-weighted imaging; CE-T1WI, contrast-
enhanced T1-weighted images.
e308 Z. Huang et al. / Clinical Radiology 79 (2024) e305ee316
3, and 5 mm (VOIperitumour 1mm, VOIperitumour 3mm, and to allow model fitting. The first selected features were
VOIperitumour 5mm), excluding the tumour region. Addition- again screened using the recursive feature elimination
ally, the MRI images of 20 patients were selected randomly (RFE) algorithm. The optimal radiomics parameters were
for re-segmentation 1 month later. Meanwhile, the VOIs in selected by the logistic regression (LR) with fivefold cross-
20 cases were re-delineated by another radiologist with 11 validation in the training cohort and were applied to the
years of experience in neurology. test cohort.
Image pre-processing was performed before feature For each model, the discrimination performance was
extraction. First, the images were resampled to a voxel size evaluated by the area under the receiver operating
of 3 3 3 mm3 using B-spline interpolation to address characteristic curve (ROC AUC) and other diagnostic
inconsistent spatial resolutions. Following the spatial indices, such as accuracy, sensitivity, and specificity.
resampling, intensity values of the MRI were normalised Moreover, the AUC values of different models were
and discretised using Z-score standardisation and five-bins compared using the DeLong test. Additionally, the diag-
histogram discretisation, respectively. nostic performance of models was then quantified by
Radiomics features were extracted from each VOI of diagnostic integrated discrimination improvement (IDI).
every single sequence using the FeAture Explorer soft- The calibration curve and the HosmereLemeshow test
ware (version 0.4.2). Each set of 107 radiomics features were applied to describe the agreement between radio-
retrieved from the original images, including the first- mics model prediction and actual EGFR mutation.
order intensity,18 three-dimensional (3D) shape fea- Moreover, decision curve analysis (DCA) and clinical
tures,14 and textural features (grey-level co-occurrence impact curve (CIC) were used to assess the clinical use-
matrix [24], grey-level size zone matrix [16], grey-level fulness and effectiveness of the radiomics model. Finally,
run length matrix [16], grey-level dependence matrix the radiomics score (R-score) predicting the probability
[14], and neighbourhood grey-tone difference matrix [5]) of EGFR mutation was calculated for each patient via a
were obtained. Z-score normalisation of the extracted linear combination of the selected features by their
radiomics features was performed before feature selec- respective coefficients, which was defined as the radio-
tion. If the feature pair’s Pearson correlation coefficient mics signature. In addition, EGFR mutational probabilities
(PCC) value was more significant than 0.99, one of them in the training and test cohorts were detected by the
was removed enough for adequate dimension reduction ManneWhitney U-test.
Z. Huang et al. / Clinical Radiology 79 (2024) e305ee316 e309
The detailed results of the 25 single-sequence radiomics In the univariate and multivariate analyses of the training
models based on different regions using the RFE-LR algo- cohort, male gender (OR¼0.214, 95% CI: 0.067e0.686,
rithm are presented in Table 3 and Fig 3. The five radiomics p¼0.009) and R-score (cut-off value 0.583) (OR¼10.155, 95%
features derived from the FLAIR radiomics model in CI: 3.541e29.121, p¼0.000) were identified as independent
VOIperitumour 1mm (radiomics model 1) showed good perfor- risk factors for EGFR mutation in the fusion model (Table 4).
mance, with a cut-off value of 0.576. In the training cohort, The AUCs (95% CI) of 0.762 (0.642e0.881) in the training
the sensitivity, specificity, and accuracy were 0.711, 0.780, cohort with a cut-off of 0.339 and 0.725 (0.541e0.909) in the
and 0.744, respectively; in the test cohort, the sensitivity, test cohort, with the sensitivity of 0.889 and 0.895, speci-
specificity, and accuracy were 0.737, 0.778, and 0.757, ficity of 0.634 and 0.556, and accuracy of 0.767 and 0.730,
respectively. The ROC curve showed AUC values of 0.746 respectively.
(95% confidence interval [CI]: 0.616e0.875, p¼0.000) and Fig. 4d provides a diagnostic nomogram. The nomogram
0.757 (95% CI: 0.562e0.952, p¼0.001) in the training and score (the optimal cut-off value of 0.339) in the training
e310
Table 2
The comparison of clinical and radiological characteristics between the training cohort and test cohort.
Characteristics Whole cohort Training cohort Test cohort Training cohort (n¼86) p-Value Test cohort (n¼37) p-Value
(n¼123) No (%) (n¼86) No (%) (n¼37) No. (%)
EGFR mutation EGFR wild-type EGFR mutation EGFR wild-type
(n¼45) No (%) (n¼41) No (%) (n¼19) No (%) (n¼18) No (%)
Age, p-value 0.111 0.643 0.737
56 years 63 (51.2) 40 (46.5) 23 (62.2) 22 (48.9) 18 (43.9) 11 (57.9) 12 (66.7)
>56 years 60 (48.8) 46 (53.5) 14 (37.8) 23 (51.1) 23 (56.1) 8 (42.1) 6 (33.3)
Gender, p-value 0.091 0.003 0.330
Female 46 (37.4) 28 (32.6) 18 (48.6) 21 (46.7) 7 (17.1) 11 (57.9) 7 (38.9)
Male 77 (62.6) 58 (67.4) 19 (51.4) 24 (53.3) 34 (82.9) 8 (42.1) 11 (61.1)
Smoking history, p-value 0.827 0.220 0.313
Never smoker 78 (63.4) 54 (62.8) 24 (64.9) 31 (68.9) 23 (56.1) 14 (73.7) 10 (55.6)
Active and former smoker 45 (36.6) 32 (37.2) 13 (35.1) 14 (31.1) 18 (43.9) 5 (26.3) 8 (44.4)
VOI, volume of interest; AUC, area under the curve; FLAIR, fluid-attenuated inversion recovery; T1WI, T1-weighted imaging; T2WI, T2-weighted imaging; RFE, recursive feature elimination; LR, logistic
regression.
e311
e312 Z. Huang et al. / Clinical Radiology 79 (2024) e305ee316
Figure 3 The heat map of AUCs in a single-sequence radiomics model based on five regions using the RFE-LR algorithm. PCC, pearson correlation
coefficient; RFE, recursive feature elimination; LR, logistic regression.
and test cohorts also demonstrated diagnostic perfor- The fusion model (AUC ¼ 0.814; Electronic Supple-
mance for EGFR mutation status by the ManneWhitney U mentary Material Fig. S2a) was better than radiomics
test (Fig 4e). model 2 (AUC ¼ 0.756) in the training cohort, and this
Z. Huang et al. / Clinical Radiology 79 (2024) e305ee316 e313
Figure 4 (aec) The optimal radiomics model based on the peritumour region of 1 mm, which was a combination of nine features selected from
the FLAIR sequence. (a) Selected features in the optimal radiomics model and corresponding coefficients. (b) EGFR nomogram. (c) Distribution of
R-scores between EGFR mutation and EGFR wild-type in the training and test cohorts. (d,e) The fusion model of EGFR mutation status. (d) EGFR
nomogram. (e) Violin and box plots of EGFR mutation probabilities in the training and test cohorts.
Table 4
Univariate and multivariate analyses of characteristics related to EGFR mutation status.
EGFR, epidermal growth factor receptor; OR, odds ratio; CI, confidence interval.
mm peritumoural area. The results of the survey by Bau- on VOIperitumour 1mm was not the same as the findings of
mert et al.13 showed that the highest percentage of infil- Wang et al.,17 who found the excellent performance of
tration was present in NSCLC (70%). There were several VOIentire tumour. This disagreement may be related to the
infiltration sites of NSCLC metastasis, but NSCLC showed a influence of EGFR mutation subtypes on radiomics features
maximum infiltration depth of <1 mm. This finding was of BMs. Although both studies retrospectively analysed lung
consistent with the observations of Neves et al. showing an adenocarcinoma patients with EGFR mutations who were
infiltration zone of 1 mm for five of 24 BMs from NSCLC.21 initially diagnosed with BMs, patients with different EGFR
Among constructed multiple VOI models, the FLAIR radio- mutation subtypes may have close relationship with
mics model obtained from the VOIperitumour 1mm (radiomics different types of BMs, consistent with literature reported
model 2) was superior to other radiomics models in pre- by Sekine et al.11 Additionally, the optimal radiomics model
dicting EGFR mutation status. The AUCs of radiomics model was established based on the VOIperitumour 1mm, which was a
2 were 0.756 in the training cohort and 0.785 in the test combination of nine features selected from the FLAIR
cohort. It was suggested that a peritumoural region of 1 mm sequence, including one first-order feature, five textural
might be associated with high tumour aggressiveness, features and three shape features. Previous researchers
coinciding with previous studies.13,22 In the present revealed that NSCLC patients with deletion of exon 19 had
research, the FLAIR radiomics model outperformed other miliary BMs, which presented more multiple and smaller
single-sequence radiomics models, in agreement with the metastatic lesions with smaller peritumoural oedema than
findings of Wang et al.,17 a FLAIR radiomics model yielded did those with EGFR wild-type.11,25e27 This finding was
AUCs of 0.987 and 0.871 in the training and test cohorts, similar to the results of the present study that the FLAIR_-
respectively; however, the optimal radiomics model based VOIperitumour 1mm_original_shape_SurfaceVolumeRatio with
Z. Huang et al. / Clinical Radiology 79 (2024) e305ee316 e315
a coefficient of -1.34 (p¼0.031) may reflect the size of analysis. This work was supported by Fujian Provincial
tumour. Health Technology Project (grant no.: 2020QNA090).
Goodman et al.13 described that the pattern of
enhancement on CE-T1WI (homogeneous, heterogeneous, Appendix A. Supplementary data
or ring) or percent necrosis of BMs, were effective pre-
dictors to predict freedom from progression (FFP) of BMs,
Supplementary data to this article can be found online at
and the lowest FFP was more likely to occur in ring-
https://doi.org/10.1016/j.crad.2023.10.022.
enhancing lesions. In the present study, only enhance-
ment patterns and necrosis of brain metastases were
significantly associated with EGFR mutations in the training References
cohort according to univariate analysis; however, no sta-
1. Imyanitov EN, Iyevleva AG, Levchenko EV. Molecular testing and tar-
tistical significance was observed in two radiological pre- geted therapy for non-small cell lung cancer: current status and per-
dictors at multivariate analysis. The enhancement pattern, spectives. Crit Rev Oncol Hematol 2021;157:103194.
or necrotic volume in patients with BMs, depends on the 2. Bousquet Mur E, Bernardo S, Papon L, et al. Notch inhibition overcomes
pathological type of the primary site. Unlike the study by resistance to tyrosine kinase inhibitors in EGFR-driven lung adenocar-
cinoma. J Clin Invest 2020;130(2):612e24.
Goodman et al.13 in which BMs originated from various
3. Midha A, Dearden S, McCormack R. EGFR mutation incidence in non-
primary tumours, such as lung (36%), melanoma (31%), small-cell lung cancer of adenocarcinoma histology: a systematic re-
breast 80 (15%), kidney (7%), colorectal (4%), and other un- view and global map by ethnicity (mutMapII). Am J Cancer Res
known primaries (7%), the present patients with BMs only 2015;5(9):2892e911.
derived from lung adenocarcinoma. Moreover, univariate 4. Jimeno A, Hidalgo M. Blockade of epidermal growth factor receptor
(EGFR) activity. Crit Rev Oncol Hematol 2005;53(3):179e92.
and multivariate analyses supported gender as an inde- 5. Imai H, Kaira K, Minato K. Clinical significance of post-progression sur-
pendent risk factor to predict the EGFR mutation status in vival in lung cancer. Thorac Cancer 2017;8(5):379e86.
the present study. It was consistent with the findings of Liu 6. Lamba N, Wen PY, Aizer AA. Epidemiology of brain metastases and
et al.,28 who reported that gender, smoking status, histo- leptomeningeal disease. Neuro Oncol 2021;23(9):1447e56.
7. Tomiyama N, Yasuhara Y, Nakajima Y, et al. CT-guided needle biopsy of
logical subtype, and pathological grade were significant
lung lesions: a survey of severe complication based on 9783 biopsies in
predictors of EGFR mutations in peripheral lung adenocar- Japan. Eur J Radiol 2006;59(1):60e4.
cinoma. Notably, they found a higher rate of EGFR muta- 8. Park YW, An C, Lee J, et al. Diffusion tensor and postcontrast T1-
tions in female patients than in male patients. The weighted imaging radiomics to differentiate the epidermal growth
discrepancy in research findings could be attributed to factor receptor mutation status of brain metastases from non-small cell
lung cancer. Neuroradiol 2021;63(3):343e52.
differences in sample sizes; the present study had a smaller
9. Chen BT, Jin T, Ye N, et al. Radiomic prediction of mutation status based
sample size. on MR imaging of lung cancer brain metastases. Magn Reson Imaging
Several constraints limit this study. It is based on a 2020;69:49e56.
retrospective design from a single centre with a small 10. Hu N, Wang G, Wu YH, et al. LDA-SVM-based EGFR mutation model for
sample size, and internal validation was performed using NSCLC brain metastases: an observational study. Medicine (Baltimore)
2015;94(5):e375.
fivefold cross-validation in the training cohort. To address 11. Sekine A, Kato T, Hagiwara E, et al. Metastatic brain tumours from non-
these limitations and improve the model, utilising a larger small cell lung cancer with EGFR mutations: distinguishing influence of
prospective dataset and conducting external validation is exon 19 deletion on radiographic features. Lung Cancer 2012;77(1):
essential. Additionally, exploring various EGFR mutation 64e9.
12. Takano K, Kinoshita M, Takagaki M, et al. Different spatial distributions
subtypes was not possible due to limited sample availability.
of brain metastases from lung cancer by histological subtype and mu-
In conclusion, the FLAIR radiomics signature in VOIper- tation status of epidermal growth factor receptor. Neuro Oncol
itumour 1mm shows promise as a potential biomarker for 2016;18(5):716e24.
predicting EGFR mutation status in lung adenocarcinoma 13. Goodman KA, Sneed PK, McDermott MW, et al. Relationship between
patients with BMs. It can significantly assist clinicians in pattern of enhancement and local control of brain metastases after
radiosurgery. Int J Radiat Oncol Biol Phys 2001;50(1):139e46.
tailoring treatment strategies based on individual patient 14. Zakaria R, Das K, Radon M, et al. Diffusion-weighted MRI characteristics
characteristics and conditions. of the cerebral metastasis to brain boundary predicts patient outcomes.
BMC Med Imaging 2014;14:26.
Conflict of interest 15. Qi Y, Zhao T, Han M. The application of radiomics in predicting gene
mutations in cancer. Eur Radiol 2022;32(6):4014e24.
16. Gillies RJ, Kinahan PE, Hricak H. Radiomics: images are more than pic-
The authors declare the following financial interests/ tures, they are data. Radiology 2016;278(2):563e77.
personal relationships which may be considered as poten- 17. Wang G, Wang B, Wang Z, et al. Radiomics signature of brain metastasis:
tial competing interests: prediction of EGFR mutation status. Eur Radiol 2021;31(7):4538e47.
18. Zheng L, Xie H, Luo X, et al. Radiomic signatures for predicting EGFR
Zhenhuan Huang reports financial support was provided
mutation status in lung cancer brain metastases. Front Oncol 2022;12:
by Fujian provincial health technology project. 931812.
19. Li Y, Lv X, Wang B, et al. Predicting EGFR T790M mutation in brain
Acknowledgements metastases using multisequence MRI-based radiomics signature. Acad
Radiol 2023;30(9):1887e95.
20. Baumert BG, Rutten I, Dehing-Oberije C, et al. A pathology-based sub-
Thea authors thank Lanmei Gao, Qiuyuan Yue, Chuan strate for target definition in radiosurgery of brain metastases. Int J
Yan, and Hanting Dai for their assistance with the radiomics Radiat Oncol Biol Phys 2006;66(1):187e94.
e316 Z. Huang et al. / Clinical Radiology 79 (2024) e305ee316
21. Neves S, Mazal PR, Wanschitz J, et al. Pseudogliomatous growth pattern 25. Ogawa M, Kurahashi K, Ebina A, et al. Miliary brain metastasis pre-
of anaplastic small cell carcinomas metastatic to the brain. Clin Neuro- senting with dementia: progression pattern of cancer metastases in the
pathol 2001;20(1):38e42. cerebral cortex. Neuropathology 2007;27(4):390e5.
22. Matsuo T, Shibata S, Yasunaga A, et al. Dose optimization and indication 26. Ruppert AM, Stankoff B, Lavole A, et al. Miliary brain metastases in lung
of Linac radiosurgery for brain metastases. Int J Radiat Oncol Biol Phys cancer. J Clin Oncol 2010;28(34):e714e6.
1999;45(4):931e9. 27. Iguchi Y, Mano K, Goto Y, et al. Miliary brain metastases from adeno-
23. Spanberger T, Berghoff AS, Dinhof C, et al. Extent of peritumoural brain carcinoma of the lung: MR imaging findings with clinical and post-
oedema correlates with prognosis, tumoural growth pattern, HIF1a mortem histopathologic correlation. Neuroradiol 2007;49(1):35e9.
expression and angiogenic activity in patients with single brain me- 28. Liu Y, Kim J, Balagurunathan Y, et al. Radiomic features are associated
tastases. Clin Exp Metastasis 2013;30(4):357e68. with EGFR mutation status in lung adenocarcinomas. Clin Lung Cancer
24. Nowosielski M, Goebel G, Iglseder S, et al. ADC textural features in pa- 2016;17(5):441e448.e446.
tients with single brain metastases improve clinical risk models. Clin Exp
Metastasis 2022;39(3):459e66.