Development and validation of a CT-based radiomics model for differentiating pneumonia-like primary pulmonary lymphoma from infectious pneumonia: A multicenter study : Chinese Medical Journal

Secondary Logo

Journal Logo

Original Article

Development and validation of a CT-based radiomics model for differentiating pneumonia-like primary pulmonary lymphoma from infectious pneumonia: A multicenter study

Yu, Xinxin1,2,; Kang, Bing2,; Nie, Pei3; Deng, Yan4; Liu, Zixin5; Mao, Ning6; An, Yahui7; Xu, Jingxu7; Huang, Chencui7; Huang, Yong8; Zhang, Yonggao9; Hou, Yang10; Zhang, Longjiang11; Sun, Zhanguo12; Zhu, Baosen1,2; Shi, Rongchao1,2; Zhang, Shuai2; Sun, Cong2,; Wang, Ximing1,

Editor(s): Gao, Ting; Hao, Xiuyuan

Author Information
Chinese Medical Journal 136(10):p 1188-1197, May 20, 2023. | DOI: 10.1097/CM9.0000000000002671

Abstract

Introduction

Primary pulmonary lymphoma (PPL) is defined as malignant monoclonal lymphoid proliferation within one or both lungs (parenchyma and/or bronchi), in a patient with no clinical, pathological, or radiographic evidence of extra-thoracic lymphoma at the time of diagnosis or during the subsequent 3 months.[1,2] PPL is a rare malignant tumor, accounting for 0.5–1% of all primary pulmonary neoplasms.[3,4] Nearly half of reported cases of PPL have been asymptomatic at the time of diagnosis, whereas in the others, the presentation has been associated with non-specific symptoms, including pulmonary findings such as cough, hemoptysis, dyspnea, or chest pain, as well as systemic symptoms like fever, weight loss, and discomfort.[5–7] In asymptomatic patients, incidental radiological findings might be the first assessment. Therefore, PPL is usually highly misdiagnosed in clinical practice.[8]

Computed tomography (CT) is a well-established modality routinely used for the initial diagnostic of lung lesion in clinical practice. Several studies have been conducted to describe the CT appearance of PPL, which showed a variety of findings, including single or multiple nodules or masses with uniform density or pneumonia-like consolidations.[2,9–11] The most frequent radiological manifestation of PPL is pneumonia-like consolidation with air bronchograms, which was commonly misdiagnosed as infectious pneumonia, leading to delayed treatment.[5,12] The mainstay of therapy for infectious pneumonia is antibiotics,[13] which is useless for those patients with PPL. Practitioners need to consider the diverse causes of a pneumonia-like syndrome before empirically prescribing antimicrobial therapy.[14] Considering the conspicuous difference in treatment, the accurate distinction between pneumonia-like PPL and infectious pneumonia is of great value, which however remains challenging due to the similar clinical and image characteristics.

As a non-invasive and promising technique, radiomics can allow high-throughput quantitative statistic features to be extracted from routine medical imaging, leading to the transformation of images into mineable data.[15,16] Successful applications of radiomics in pulmonary lesion have been reported in detection of lung cancer, prediction of histology and subtype, assessment of treatment effect, and prediction of survival.[17–20] However, to the best of our knowledge, no study has evaluated radiomics for its ability to differentiate pneumonia-like PPL from infectious pneumonia.

This study aimed to develop and validate a CT-based radiomics model for the differentiation pneumonia-like PLL from infectious pneumonia, and to help clinicians to set the right therapy.

Methods

Patients

The Institute Review Board of Shandong Provincial Hospital approved this multicenter retrospective study (No.2022-386). The requirement for obtaining informed consent was waived. The study was conducted in accordance with the Declaration of Helsinki (as revised in 2000).

Patients with pathological diagnosis of PPL, identified by searching the pathology database from 12 medical centers between January 2007 and December 2021, were enrolled. Inclusion criteria were as follows: (a) pathological diagnosis with PPL; (b) no evidence of extra-thoracic lymphoma at the time of primary diagnosis or within 3 months since onset; and (c) patients with complete clinic data. Exclusion criteria were as follows: (a) an interval longer than 4 weeks between CT and surgery; (b) image quality unsatisfactory for analysis; (c) patients received chemotherapy or radiotherapy before CT; and (d) patients without pneumonia-like lesion. The pneumonia-like lesion was defined as single or multiple patchy consolidations along lobe or segment of unilateral or bilateral lungs.[11] Finally, a total of 79 pneumonia-like PPL patients (mean age, 58 ± 12 years; 43 men) were included in this study (Center 1, Shandong Provincial Hospital [n = 4]; Center 2, Shandong Cancer Hospital and Institute [n = 10]; Center 3, The General Hospital of the People's Liberation Army [n = 13]; Center 4, The First Affiliated Hospital of Zhengzhou University [n = 10]; Center 5, Jinling Hospital [n = 7]; Center 6, Affiliated Hospital of Jining Medical University [n = 4]; Center 7, Shengjing Hospital of China Medical University [n = 8]; Center 8, Qilu Hospital of Shandong University [n = 5]; Center 9, Yantai Yuhuangding Hospital [n = 5]; Center 10, The Affiliated Hospital of Qingdao University [n = 7]; Center 11, Shandong Provincial Qianfoshan Hospital [n = 2]; Center 12, Linyi People's Hospital [n = 4]). At the same time, 176 patients (mean age, 60 ± 16 years; 106 men) with infectious pneumonia between January 2016 and December 2021 were enrolled according to the following inclusion criteria: (a) pathological or clinical confirmation of inflammatory lesion; and (b) patients with complete clinic data. Exclusion criteria were as follows: (a) an interval longer than 4 weeks between CT and treatment; (b) image quality unsatisfactory for analysis; (c) patients with anti-inflammatory therapy prior to the initial chest CT scan; and (d) patients without obvious patchy consolidations on CT. Apart from proven by pathologically (n = 48), infectious pneumonia was clinically diagnosed by fever or cough, confirmation of pathogens on sputum examination, which was diagnosed by gram stain and bacterial culture, or resolved parenchymal consolidations on the follow-up imaging after treatment with antibiotics.

If a patient had multiple lesions, then the single pathologically proven lesion (for pneumonia-like PPL) or the largest lesion (for infectious pneumonia) was analyzed per patient.

Patients from center 1 to center 7 were assigned to the training or validation cohort, and the remaining patients from other centers were used as the external test cohort [Figure 1].

F1
Figure 1:
Flowchart for selecting the study population of pneumonia-like PPL and infectious pneumonia. CT: Computed tomography; PPL: Primary pulmonary lymphoma.

For each patient, clinical features that possibly contributed to differentiate pneumonia-like PPL from infectious pneumonia were initially included. The clinical features consisted of demographic information: age, sex, and smoking history; clinical manifestations: fever, weight loss, cough, productive cough, chest pain, and blood in sputum; and laboratory parameters: hemoglobin, white blood cell (WBC), and C-reactive protein (CRP). Clinical manifestations were symptoms at the time of admission, and laboratory parameters were carried out by blood test within three days after admission.

CT scanning and CT features evaluation

Non-enhanced CT was performed using one of the following nine multi-detector CT scanners: Somatom Force, Somatom Definition Flash, and Somatom Sensation 64 (Siemens Healthcare, Erlangen, Germany); Revolution, Discovery 750, Optima 620, and Brightspeed (GE Healthcare, Milwaukee, Wisconsin, USA); and Brilliance iCT 256 and Brilliance 16 (Philips, Cleveland, Ohio, USA). Detailed acquisition and reconstruction parameters are shown in Supplementary Table 1, https://links.lww.com/CM9/B516.

Two radiologists (X.W. and X.Y.) who were blinded to the clinical and pathologic results evaluated CT features. They assessed the following imaging features of lesions by consensus: margin, attenuation, maximum diameter, air bronchogram, irregular cavitation, halo sign, thickening of bronchiovascular bundles, pleural indentation sign, pleural effusion, and lymphadenopathy [Figure 2].

F2
Figure 2:
CT features of pneumonia-like PPL. (A) Lung consolidation with subpleural ill-defined ground-glass halo sign (white arrowheads) in the left upper lobe of a 61-year-old male patient with MALT lymphoma. (B) The ill-marginated patchy consolidation with irregular cavitations (red arrow) and interlobular septa involvement in right upper lobe of a 69-year-old male with MALT lymphoma. (C) Bilateral peribronchovascular patchy consolidations with irregular air bronchogram sign (red arrows) and pleural indentation sign (white arrowheads) in a 58-year-old male patient with small B-cell lymphoma. (D) Massive consolidations with air bronchogram sign, bronchiectasia (red arrow), and pleural effusion (white arrowheads) in bilateral lungs in a 63-year-old female patient with MALT lymphoma. MALT: Mucosa-associated lymphoid tissue; PPL: Primary pulmonary lymphoma.

Development of clinical factor model

Univariate analysis was used to assess the difference of clinicoradiological variables (including clinical and CT features) between pneumonia-like PPL and infectious pneumonia in the training cohort. Then, variables with P <0.05 in univariate analysis were applied to a multivariate logistic regression analysis to elucidate the independent factors. Meanwhile, the clinical factor model was built based on these independent factors. Odds ratios (OR) with 95% confidence intervals (CIs) were calculated for each risk factor.

CT visual assessment

Two radiologists (X.W. [Reader 1] and B.K. [Reader 2], with 20 years and six years of experience, respectively, in pulmonary imaging) retrospectively and independently reviewed the pre-treatment CT scans from the external test cohort. Blinded to the clinical data and pathological diagnosis, they were encouraged to decide whether each lesion was either PPL or infectious pneumonia using a 5-point scale (0, definite infectious pneumonia; 1, probable infectious pneumonia; 2, equivocal; 3, probable PPL; and 4, definite PPL). The accuracy, sensitivity, and specificity of the readers were calculated by converting the 5-point scale into a binary class. A score of 0 or 1 was regarded as infectious pneumonia, and a score of 2–4 was regarded as pneumonia-like PPL.

Three-dimensional segmentation of lesions and radiomics features extraction

Three-dimensional segmentation of the region of interest (ROI) was semiautomatically segmented using attenuation-based thresholding, and then manually verified and corrected slice-by-slice by one radiologist (X.Y.) on axial CT images in the lung window, within the edge of the lesion, while avoiding adjacent normal lung tissue. To assess the robustness of segmentation, 30 randomly selected lesions (containing 10 pneumonia-like PPL and 20 infectious pneumonia) were contoured twice by the same radiologist (X.Y.) in a 4-week period, as well as by another radiologist (B.K.).

In order to decrease the variability of the radiomics feature, gray-level discretization and image resampling were performed prior to feature extraction, due to the images having been acquired from various scanners with different parameters. All CT images were resampled to a pixel spacing of 1.0 mm × 1.0 mm × 1.0 mm using sitkBSpline methods and discretized with a bin width of 25 HU. The Dr. Wise Multimodal Research Platform (V1.6.2, https://keyan.deepwise.com) was used to extract a total of 1743 radiomics features, including first-order features, texture features, and shape features from the original images and transformed images. To guarantee the repeatability of the results, z-score normalization was then performed as preprocessing step.

Development of radiomics model

We devised a three-step procedure for high-dimensional radiomics feature selection. First, the inter- and intra-class correlation coefficients (ICCs) were calculated to explore the stability of radiomics features, and features with an ICC <0.75 were excluded. Furthermore, as assessing repeatability and reproducibility by calculation of ICC might not be sufficient, we also calculated the coefficient of variation (CV), another precision test commonly used for repeatability and reproducibility experiments based on intra-subject variability. Second, a one-way analysis of variance (ANOVA) was performed on the remaining features to select statistically significant features. Finally, the selected features were then enrolled into a least absolute shrinkage and selection operator (LASSO) logistic regression algorithm with five-fold cross-validation to identify the most valuable features (with non-zero coefficients).

Statistical analysis

The statistical analysis was conducted using the SPSS 22.0 software (IBM Corp. SPSS Inc., Chicago, USA) and MedCalc 20 software (MedCalc Software, Ostend, Belgium). The unpaired t-test was used for continuous variables, and the chi-squared test or Fisher's exact test was used for categorical variables. Multivariate logistic regression analysis was used to select the independent factors and build the clinical factor model. The performances of the clinical factor model, two readers, and radiomics model were assessed by area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and accuracy. The optimal cut-off value was determined by maximizing Youden's index. The AUC of the radiomics model was compared with those of the two readers and the clinical factor model using the DeLong method.[21]P <0.05 was indicative of statistical significance.

Results

Patient characteristics

A total of 255 patients were included for the training cohort (n = 144; mean age, 60 ± 15 years; 89 men), validation cohort (n = 38; mean age, 54 ± 17 years; 20 men), and external test cohort (n = 73; mean age, 61 ± 12 years; 40 men) in this study. The baseline characteristics of all patients are summarized in Table 1. The rates of pneumonia-like PPL were 30.6% (44 of 144), 31.6% (12 of 38), and 31.5% (23 of 73) in the training, validation, and external test cohorts, respectively, whereas no statistically significant difference was found (P =0.986).

Table 1 - Comparisons of baseline characteristics between pneumonia-like PPL and infectious pneumonia.
Characteristics Training cohort P-values Validation cohort P-values External test cohort P-values

PPL

(n = 44)

Pneumonia

(n = 100)

PPL

(n = 12)

Pneumonia

(n = 26)

PPL

(n = 23)

Pneumonia

(n = 50)

Patient demographics
Age (years) 60 ± 11 60 ± 16 0.984 56 ± 12 52 ± 19 0.591 56 ± 13 63 ± 11 0.028
Men 26 (59.1) 63 (63.0) 0.657 6 (50.0) 14 (53.8) 0.825 11 (47.8) 29 (58.0) 0.658
Smokers 16 (36.4) 43 (43.0) 0.456 3 (25.0) 4 (15.4) 0.656 5 (21.7) 22 (44.0) 0.067
Clinical manifestations
Fever 4 (9.1) 48 (48.0) <0.001 1 (8.3) 12 (46.2) 0.030 5 (21.7) 31 (62.0) 0.001
Weight loss 3 (6.8) 8 (8.0) >0.999 0 4 (15.4) 0.287 3 (13.0) 6 (12.0) >0.999
Cough 30 (68.2) 81 (81.0) 0.092 5 (41.7) 17 (65.4) 0.169 10 (43.5) 32 (64.0) 0.099
Productive cough 22 (50.0) 77 (77.0) 0.001 3 (25.0) 15 (57.7) 0.086 8 (34.8) 30 (60.0) 0.045
Blood in sputum 5 (11.4) 4 (4.0) 0.132 1 (8.3) 2 (7.7) >0.999 1 (4.3) 7 (14.0) 0.422
Chest pain 12 (27.3) 10 (10.0) 0.008 0 2 (7.7) >0.999 2 (8.7) 15 (30.0) 0.072
Laboratory parameters
Hemoglobin 0.456 >0.999 0.093
Normal 28 (63.6) 57 (57.0) 9 (75.0) 18 (69.2) 18 (78.3) 29 (58.0)
Decreased 16 (36.4) 43 (43.0) 3 (25.0) 8 (30.8) 5 (21.7) 21 (42.0)
WBC 0.088 0.472 0.062
Normal 36 (81.8) 68 (68.0) 9 (75.0) 15 (57.7) 19 (82.6) 29 (58.0)
Elevated 8 (18.2) 32 (32.0) 3 (25.0) 11 (42.3) 4 (17.4) 21 (42.0)
CRP 0.001 0.013 <0.001
Normal 28 (63.6) 34 (34.0) 10 (83.3) 9 (34.6) 17 (73.9) 7 (14.0)
Elevated 16 (36.4) 66 (66.0) 2 (16.7) 17 (65.4) 6 (26.1) 43 (86.0)
CT features
Attenuation 37.2 ± 10.2 32.8 ± 5.4 0.001 36.7 ± 9.7 31.7 ± 4.6 0.038 35.9 ± 9.8 32.7 ± 4.8 0.064
Diameter (cm) 7.3 ± 3.2 7.5 ± 2.5 0.564 6.5 ± 3.7 7.0 ± 1.9 0.580 6.5 ± 3.3 8.0 ± 3.3 0.078
Margin 0.002 0.068 0.030
Well-defined 14 (31.8) 11 (11.0) 6 (50.0) 5 (19.2) 9 (39.1) 8 (16.0)
Poorly defined 30 (68.2) 89 (89.0) 6 (50.0) 21 (80.8) 14 (60.9) 42 (84.0)
Air bronchogram 0.011 0.009 0.903
Absent 4 (9.1) 19 (19.0) 2 (16.7) 1 (3.8) 2 (8.7) 6 (12.0)
Regular 11 (25.0) 42 (42.0) 3 (25.0) 20 (76.9) 11 (47.8) 24 (48.0)
Irregular 29 (65.9) 39 (39.0) 7 (58.3) 5 (19.2) 10 (43.5) 20 (40.0)
Irregular cavitation <0.001 0.235 0.760
Absent 16 (36.4) 68 (68.0) 7 (58.3) 21 (80.8) 12 (52.2) 28 (56.0)
Present 28 (63.6) 32 (32.0) 5 (41.7) 5 (19.2) 11 (47.8) 22 (44.0)
Halo sign 0.001 >0.999 0.092
Absent 22 (50.0) 79 (79.0) 9 (75.0) 19 (73.1) 15 (65.2) 22 (44.0)
Present 22 (50.0) 21 (21.0) 3 (25.0) 7 (26.9) 8 (34.8) 28 (56.0)
Thickening of bronchiovascular bundles 0.304 0.016 0.048
Absent 23 (52.3) 43 (43.0) 9 (75.0) 8 (30.8) 18 (78.3) 27 (54.0)
Present 21 (47.7) 57 (57.0) 3 (25.0) 18 (69.2) 5 (21.7) 23 (46.0)
Pleural indentation sign 0.003 0.486 0.238
Absent 13 (29.5) 56 (56.0) 9 (75.0) 16 (61.5) 14 (60.9) 23 (46.0)
Present 31 (70.5) 44 (44.0) 3 (25.0) 10 (38.5) 9 (39.1) 27 (54.0)
Pleural effusion <0.001 0.003 0.012
Absent 30 (68.2) 34 (34.0) 10 (83.3) 8 (30.8) 19 (82.6) 26 (52.0)
Present 14 (31.8) 66 (66.0) 2 (16.7) 18 (69.2) 4 (17.4) 24 (48.0)
Lymphadenopathy 0.387 0.068 >0.999
Absent 23 (52.3) 60 (60.0) 6 (50.0) 21 (80.8) 15 (65.2) 34 (68.0)
Present 21 (47.7) 40 (40.0) 6 (50.0) 5 (19.2) 8 (34.8) 16 (32.0)
Data are presented as n (%) or mean ± standard deviation. CRP: C-reactive protein; CT: Computed tomography; PPL: Primary pulmonary lymphoma; WBC: White blood cell.

Development of clinical factor model

To identify the potential clinicoradiological parameters used for clinical model development, univariate analysis revealed that fever, productive cough, chest pain, CRP, CT attenuation, margin, irregular cavitation, halo sign, pleural indentation sign, and pleural effusion were tightly correlated to pneumonia-like PPL (P <0.05). The multivariate logistic regression analysis revealed that fever (OR = 0.27; 95% CI: 0.08–0.96; P =0.043), productive cough (OR = 0.24; 95% CI: 0.08–0.68; P =0.008), and irregular cavitation (OR = 2.88; 95% CI: 1.03–8.11; P =0.045) remained independent predictors of pneumonia-like PPL [Table 2]. A clinical factor model was developed based on the above independent predictors.

Table 2 - Univariate and multivariate logistic regression analysis of clinicoradiological characteristics for differentiating pneumonia-like PPL from infectious pneumonia.
Characteristics Univariable analysis Multivariable analysis
ORs (95% CI) P-values ORs (95% CI) P-values
Age 1.00 (0.98–1.03) 0.984
Men 0.85 (0.41–1.75) 0.657
Smoke 0.76 (0.37–1.57) 0.456
Fever 0.11 (0.04–0.33) <0.001 0.27 (0.08–0.96) 0.043
Weight loss 0.84 (0.21–3.34) 0.806
Cough 0.50 (0.22–1.13) 0.095
Productive cough 0.30 (0.14–0.63) 0.002 0.24 (0.08–0.68) 0.008
Blood in sputum 3.08 (0.79–12.07) 0.107
Chest pain 3.38 (1.33–8.56) 0.010 3.10 (0.90–10.69) 0.074
Hemoglobin, decreased 0.76 (0.37–1.57) 0.456
WBC, elevated 0.47 (0.20–1.13) 0.092
CRP, elevated 0.29 (0.14–0.62) 0.001 0.43 (0.15–1.17) 0.098
CT attenuation 1.09 (1.03–1.15) 0.001 1.07 (1.00–1.15) 0.062
Diameter 0.96 (0.84–1.10) 0.561
Margin, poorly defined 0.27 (0.11–0.65) 0.003 0.49 (0.15–1.59) 0.233
Air bronchogram 1.24 (0.35–4.41) 0.735
Irregular cavitation 3.72 (1.77–7.83) 0.001 2.88 (1.03–8.11) 0.045
Halo sign 3.76 (1.76–8.06) 0.001 2.70 (0.98–7.42) 0.055
Thickening of bronchiovascular bundles 0.69 (0.34–1.40) 0.305
Pleural indentation sign 3.04 (1.42–6.48) 0.004 2.70 (0.98–7.40) 0.054
Pleural effusion 0.24 (0.11–0.51) <0.001 0.41 (0.15–1.11) 0.080
Lymphadenopathy 1.37 (0.67–2.80) 0.388
CI: Confidence interval; CRP: C-reactive protein; CT: Computed tomography; OR: Odds ratio; PPL: Primary pulmonary lymphoma; WBC: White blood cell.

Development and testing of the radiomics model

To ensure stability and reproducibility of the radiomics features in segmentation, 195 ineligible features (178 features with ICCs <0.75 and 17 features with variance close to 0) were excluded. In the one-way ANOVA, a total of 955 features showed statistically significant differences between pneumonia-like PPL and infectious pneumonia. Using the LASSO regression model with five-fold cross-validation, 23 features with non-zero coefficients were finally selected [Supplementary Figure 1, https://links.lww.com/CM9/B516]. Detailed repeatability results of the selected radiomics features are shown in Supplementary Table 2, https://links.lww.com/CM9/B516. All of the features showed excellent repeatability/reproducibility (CV ≤10%). The heatmap of these 23 radiomics features showed differences between pneumonia-like PPL and infectious pneumonia [Figure 3]. The calculation formula for the radiomics model was attached in Supplementary Materials, https://links.lww.com/CM9/B516.

F3
Figure 3:
Heatmap of the 23 relevant radiomics features for differentiating between PPL and pneumonia. PPL: Primary pulmonary lymphoma.

In the training cohort, the AUC of the radiomics model was 0.95 (95% CI: 0.94–0.99). The sensitivity, specificity, and accuracy were 89% (39 of 44 patients; 95% CI: 0.83–0.94), 87% (87 of 100 patients; 95% CI: 0.82–0.92), and 88% (126 of 144 patients; 95% CI: 0.82–0.93), respectively. According to the cutoff value, determined by the highest Youden index, the AUC, sensitivity, specificity, and accuracy in the validation cohort were 0.93 (95% CI: 0.85–0.98), 83% (10 of 12 patients; 95% CI: 0.71–0.95), 81% (21 of 26 patients; 95% CI: 0.68–0.93), and 82% (31 of 38 patients; 95% CI: 0.69–0.94), respectively. In the external test cohort, the AUC, sensitivity, specificity, and accuracy were 0.94 (95% CI: 0.87–0.99), 96% (22 of 23 patients; 95% CI: 0.91–1.00), 80% (40 of 50 patients; 95% CI: 0.71–0.89), and 85% (62 of 73 patients; 95% CI: 0.77–0.93), respectively. The calibration curve of the radiomics model is presented in Figure 4.

F4
Figure 4:
Calibration curve of radiomics model in the training (A), validation (B), and external test (C) cohorts for differentiating pneumonia-like PPL from infectious pneumonia. PPL: Primary pulmonary lymphoma.

Comparison between the radiomics model and readers' interpretation

In the external test cohort, the AUCs were 0.74 (95% CI: 0.63–0.83) and 0.72 (95% CI: 0.62–0.82) for readers 1 and 2, respectively. The sensitivity, specificity, and accuracy, respectively, were 74% (17 of 23 patients; 95% CI: 0.64–0.84), 66% (33 of 50 patients; 95% CI: 0.55–0.77), and 68% (50 of 73 patients; 95% CI: 0.58–0.79) for reader 1; and 48% (11 of 23 patients; 95% CI: 0.36–0.59), 88% (44 of 50 patients; 95% CI: 0.81–0.95), and 75% (55 of 73 patients; 95% CI: 0.65–0.85) for reader 2. The diagnostic performance of the radiomics model was higher than that of reader 1 and 2 (0.94 vs. 0.74 [P = 0.006] and 0.94 vs. 0.72 [P = 0.003], respectively) in the external test cohort [Figure 5].

F5
Figure 5:
The ROC curves of the clinical factor model, the radiomics model, and the two readers' evaluation for differentiating pneumonia-like PPL from infectious pneumonia in the external test cohort. PPL: Primary pulmonary lymphoma; ROC: Receiver operating characteristic.

Comparison between the radiomics model and clinical factor model

The AUC, sensitivity, specificity, and accuracy of the clinical factor model for differentiating pneumonia-like PPL and infectious pneumonia were 0.80 (95% CI: 0.73–0.86), 77% (34 of 44 patients; 95% CI: 0.70–0.84), 69% (69 of 100 patients; 95% CI: 0.61–0.77), and 72% (103 of 144 patients; 95% CI: 0.64–0.79) in the training cohort, respectively. According to the cutoff value, determined by the highest Youden index, the AUC, sensitivity, specificity, and accuracy in the validation cohort were 0.75 (95% CI: 0.62–0.87), 75% (9 of 12 patients; 95% CI: 0.61–0.89), 62% (16 of 26 patients; 95% CI: 0.46–0.77), and 66% (25 of 38 patients; 95% CI: 0.51–0.81), respectively. The AUC, sensitivity, specificity, and accuracy in the external test cohort were 0.73 (95% CI: 0.62–0.84), 70% (16 of 23 patients; 95% CI: 0.59–0.80), 66% (33 of 50 patients; 95% CI: 0.55–0.77), and 67% (49 of 73 patients; 95% CI: 0.56–0.78), respectively.

Compared with the clinical factor model only, the radiomics model exhibited better predictive performance for distinguishing pneumonia-like PPL from infectious pneumonia in the validation (0.93 vs. 0.75, P =0.042) and external test cohort (0.94 vs. 0.73, P =0.008) [Table 3].

Table 3 - Diagnostic performance of the radiomics model, the two readers, and the clinical factor model in the external test cohort for differentiating pneumonia-like PPL from infectious pneumonia.
Parameter Radiomics model Reader 1 Reader 2 Clinical factor model
AUC 0.94 (95% CI: 0.87–0.99) 0.74 (95% CI: 0.63–0.83) 0.72 (95% CI: 0.62–0.82) 0.73 (95% CI: 0.62–0.84)
Sensitivity (%) 96 (22/23) 74 (17/23) 48 (11/23) 70 (16/23)
Specificity (%) 80 (40/50) 66 (33/50) 88 (44/50) 66 (33/50)
Accuracy (%) 85 (62/73) 68 (50/73) 75 (55/73) 67 (49/73)
AUC: Area under the receiver operating characteristic curve; CIs: Confidence intervals; PPL: Primary pulmonary lymphoma. Data in parentheses are numbers of patients. The radiomics model outperformed reader 1, reader 2, and clinical factor model (P = 0.006, 0.003, and 0.008 respectively).

Discussion

Differentiating PPL and infectious pneumonia is challenging when a pneumonia-like consolidation is encountered at CT. In this study, we developed a radiomics model for distinguishing pneumonia-like PPL from infectious pneumonia with the use of unenhanced CT. The radiomics model demonstrated excellent diagnostic performance with the external test cohort (AUC, 0.94; sensitivity, 96%; specificity, 80%; and accuracy, 85%). The diagnostic performance of the radiomics model outperformed the readers' interpretation (AUC, 0.74 [P = 0.006] and 0.72 [P = 0.003]). In addition, the radiomics model showed better diagnostic performance compared with the clinical factor model in the validation (AUC, 0.75 [P = 0.042]) and external test cohort (AUC, 0.73 [P = 0.008]).

To date, no guidelines for treatment have been established for PPL. In consideration of the indolent characteristic, many researchers recommended that observation and regular follow-up are the first choice.[22] Besides, different therapies have been used in the literature, including surgery, chemotherapy, or radiotherapy.[23–25] According to the guideline represent the view of National Institute for Health and Care Excellence (NICE), pneumonia is diagnosed in 5%–12% of adults who present to general practitioners with symptoms of lower respiratory tract infection, and approximately one-third of these are admitted to hospital, where the mortality rate is about 5%–14%.[26] For those patients with infectious pneumonia, antibiotic treatment should be offered as soon as possible after establishing the diagnosis.[13,27,28]

Even though several studies have endeavored to explore the CT characteristics of PPL, diagnosis of this disease entity remains an intractable conundrum due to its non-specific and diverse radiological abnormalities.[7,10,29,30] The imaging findings of PPL can be divided into three patterns: single nodular/mass pattern, multiple nodular/mass pattern, and pneumonia-like consolidative pattern.[11] Zhang et al[31] reviewed non-contrast enhanced CT in 62 PPL patients and revealed that the most frequent findings of PPL are consolidations, followed by nodules and masses. Bi et al[11] concluded that in consolidative forms, air bronchogram sign, irregular cavitation, a well-marginated halo sign, and thickening of bronchiovascular bundles were common features. However, these findings are not sufficient to help radiologists make differential diagnoses. In our study, irregular cavitation in the lesion was an independent factor for predicting PPL, whereas the other features showed no significant difference between PPL and pneumonia in the multiple logistic regression analysis, which means that the imaging findings of PPL and infectious pneumonia may overlap to some extent. The irregular cavitation communicating with airways might be in reality cystic bronchiectasis, which occurred more commonly in PPL. According to past research, dilated airways may result from the fact that most tumor cells spread surrounding the bronchi and manufactured abundant fibrous tissue that could generate various tractions of the bronchial wall.[11] Several studies had reported cases in which PPL presented as infectious pneumonia, and sometimes it can mimic non-resolving pneumonia.[5,29] In a retrospective study, more than half of patients (21/36) with PPL were initially misdiagnosed as having pneumonia, and received antibiotics for 1–4 weeks before hospitalization.[32] Our study also confirmed this phenomenon. The radiologists showed poor performance in the prediction of pneumonia-like PPL, with the AUC of 0.74 and 0.72 in the external test cohort. To improve the radiological diagnosis capacity, radiomics combining with machine learning is the most potential method waiting to be explored.

Radiomics has confirmed clinical value because of its potential capacity to capture high-throughput useful information and to identify subtle changes that are difficult to ascertain by visual assessment.[15] In fact, radiomics has been used in many researches for promoting diagnosis, staging, and therapy response of many diseases.[19,33] Previous studies have shown that radiomics can be used for the classification of pulmonary lesions. Wu et al[34] reported that CT-based radiomics signature was useful in diagnosing the invasiveness of lung adenocarcinoma. Especially in this era of COVID-19 pneumonia, various studies have described the performance of radiomics signature for differentiation of COVID-19[35,36] and predicting the severity.[37] In this study, we first compared the radiomics features of pneumonia-like PPL with infectious pneumonia. For the fusion radiomics model, 23 radiomics features were selected after LASSO, including 8 first order, 2 shape, and 13 texture features, which were transformed by various filters. Most radiomics features in this study were extracted from multiple frequency and spatial domains, which would be difficult to explain intuitively in the clinical scenario.[38] Among the selected radiomics features, lbp-3D-k_firstorder_Maximum, wavelet-HH_gldm_DependenceVariance and logarithm_gldm_Large Dependence High Gray Level Emphasis were the most significant and robust features associated with PPL, which reflect a lesion's intensity and textural features within the high-intensity CT voxels. First-order features describe the distribution of voxel signal intensities within ROI through commonly used metrics, in which maximum represents the maximum gray-level value in ROI, which means that pneumonia-like PPL has a higher gray-level value than infectious pneumonia. GLDM quantifies the dependence of gray-level in an image, and DependenceVariance measures the variance of dependence size with a higher value indicating greater dependence difference and more heterogeneous textures.

Several limitations exist in our study. First, the patient population was relatively small due to the strict inclusion criteria. A large-scale independent prospective multicenter study is needed to evaluate the generalizability of the results. Second, the retrospective nature might have inevitably introduced bias in population selection. The two groups in our study population were unbalanced, which might indicate a spectrum bias and might have influenced the diagnostic performance. Third, CT acquisition parameters and reconstruction techniques were not consistent due to the retrospective and multi-institutional nature of the study, which might have resulted in the variability of CT attenuation values with resultant bias for estimation of radiomic features. Finally, different observers for segmentation could have affected the stability of radiomics features. Although only features with ICCs >0.75 were kept for radiomics model construction in our study, automated, reliable, and robust boundary extraction method must be developed to facilitate the efficiency of the radiomics process.

In conclusion, a CT-based radiomics model can be useful for differentiating pneumonia-like PPL and infectious pneumonia. The diagnostic performance of radiomics model was superior to that of radiologists and the clinical factor model.

Funding

This work was supported by grants from the National Natural Science Foundation of China (Nos. 81871354 and 81571672) and the Academic Promotion Program of Shandong First Medical University (No. 2019QL023).

Conflicts of interest

None.

References

1. Cadranel J, Wislez M, Antoine M. Primary pulmonary lymphoma. Eur Respir J 2002;20: 750–762. doi: 10.1183/09031936.02.00404102.
2. Graham BB, Mathisen DJ, Mark EJ, Takvorian RW. Primary pulmonary lymphoma. Ann Thorac Surg 2005;80: 1248–1253. doi: 10.1016/j.athoracsur.2005.04.014.
3. Miller DL, Allen MS. Rare pulmonary neoplasms. Mayo Clin Proc 1993;68: 492–498. doi: 10.1016/s0025-6196(12)60199-2.
4. William J, Variakojis D, Yeldandi A, Raparia K. Lymphoproliferative neoplasms of the lung: A review. Arch Pathol Lab Med 2013;137: 382–391. doi: 10.5858/arpa.2012-0202-RA.
5. Yao D, Zhang L, Wu PL, Gu XL, Chen YF, Wang LX, et al. Clinical and misdiagnosed analysis of primary pulmonary lymphoma: A retrospective study. BMC Cancer 2018;18: 281. doi: 10.1186/s12885-018-4184-1.
6. He H, Tan F, Xue Q, Liu L, Peng Y, Bai G, et al. Clinicopathological characteristics and prognostic factors of primary pulmonary lymphoma. J Thorac Dis 2021;13: 1106–1117. doi: 10.21037/jtd-20-3159.
7. Zhang XY, Gu DM, Guo JJ, Su QQ, Chen YB. Primary pulmonary lymphoma: A retrospective analysis of 27 cases in a single tertiary hospital. Am J Med Sci 2019;357: 316–322. doi: 10.1016/j.amjms.2019.01.009.
8. Gao X, Yan P. Primary pulmonary lymphoma presenting as nodules and masses: A case report and literature review. Monaldi Arch Chest Dis 2021;91. doi: 10.4081/monaldi.2021.1259.
9. King LJ, Padley SP, Wotherspoon AC, Nicholson AG. Pulmonary MALT lymphoma: Imaging findings in 24 cases. Eur Radiol 2000;10: 1932–1938. doi: 10.1007/s003300000491.
10. Albano D, Borghesi A, Bosio G, Bertoli M, Maroldi R, Giubbini R, et al. Pulmonary mucosa-associated lymphoid tissue lymphoma: 18F-FDG PET/CT and CT findings in 28 patients. Br J Radiol 2017;90: 20170311. doi: 10.1259/bjr.20170311.
11. Bi W, Zhao S, Wu C, Gao J, Zhao S, Yang S, et al. Pulmonary mucosa-associated lymphoid tissue lymphoma: CT findings and pathological basis. J Surg Oncol 2021;123: 1336–1344. doi: 10.1002/jso.26403.
12. Cozzi D, Dini C, Mungai F, Puccini B, Rigacci L, Miele V. Primary pulmonary lymphoma: Imaging findings in 30 cases. Radiol Med 2019;124: 1262–1269. doi: 10.1007/s11547-019-01091-z.
13. Cillóniz C, Torres A, Niederman MS. Management of pneumonia in critically ill patients. BMJ 2021;375: e065871. doi: 10.1136/bmj-2021-065871.
14. Mandell LA, Wunderink RG, Waterer GW. Community-acquired pneumonia. N Engl J Med 2015;372: 293–294. doi: 10.1056/NEJMc1414306.
15. Gillies RJ, Kinahan PE, Hricak H. Radiomics: Images are more than pictures, they are data. Radiology 2016;278: 563–577. doi: 10.1148/radiol.2015151169.
16. Lambin P, Leijenaar RTH, Deist TM, Peerlings J, de Jong EEC, van Timmeren J, et al. Radiomics: The bridge between medical imaging and personalized medicine. Nat Rev Clin Oncol 2017;14: 749–762. doi: 10.1038/nrclinonc.2017.141.
17. Huang P, Park S, Yan R, Lee J, Chu LC, Lin CT, et al. Added value of computer-aided CT image features for early lung cancer diagnosis with small pulmonary nodules: A matched case-control study. Radiology 2018;286: 286–295. doi: 10.1148/radiol.2017162725.
18. Cho YH, Seo JB, Lee SM, Kim N, Yun J, Hwang JE, et al. Radiomics approach for survival prediction in chronic obstructive pulmonary disease. Eur Radiol 2021;31: 7316–7324. doi: 10.1007/s00330-021-07747-7.
19. Wang T, She Y, Yang Y, Liu X, Chen S, Zhong Y, et al. Radiomics for survival risk stratification of clinical and pathologic stage IA pure-solid non-small cell lung cancer. Radiology 2022;302: 425–434. doi: 10.1148/radiol.2021210109.
20. Chetan MR, Gleeson FV. Radiomics in predicting treatment response in non-small-cell lung cancer: Current status, challenges and future perspectives. Eur Radiol 2021;31: 1049–1058. doi: 10.1007/s00330-020-07141-9.
21. DeLong ER, DeLong DM, Clarke-Pearson DL. Comparing the areas under two or more correlated receiver operating characteristic curves: A nonparametric approach. Biometrics 1988;44: 837–845. doi: 10.2307/2531595.
22. Stefanovic A, Morgensztern D, Fong T, Lossos IS. Pulmonary marginal zone lymphoma: A single centre experience and review of the SEER database. Leuk Lymphoma 2008;49: 1311–1320. doi: 10.1080/10428190802064933.
23. Zucca E, Conconi A, Martinelli G, Bouabdallah R, Tucci A, Vitolo U, et al. Final results of the IELSG-19 randomized trial of mucosa-associated lymphoid tissue lymphoma: Improved event-free and progression-free survival with rituximab plus chlorambucil versus either chlorambucil or rituximab monotherapy. J Clin Oncol 2017;35: 1905–1912. doi: 10.1200/JCO.2016.70.6994.
24. Conconi A, Martinelli G, Thiéblemont C, Ferreri AJ, Devizzi L, Peccatori F, et al. Clinical activity of rituximab in extranodal marginal zone B-cell lymphoma of MALT type. Blood 2003;102: 2741–2745. doi: 10.1182/blood-2002-11-3496.
25. Wang L, Ye G, Liu Z, Shi L, Zhan C, Gu J, et al. Clinical characteristics, diagnosis, treatment, and prognostic factors of pulmonary mucosa-associated lymphoid tissue-derived lymphoma. Cancer Med 2019;8: 7660–7668. doi: 10.1002/cam4.2683.
26. NIH. Pneumonia in adults: Diagnosis and management. London: National Institute for Health and Care Excellence Guidelines, 2019.
27. Kallander K, Burgess DH, Qazi SA. Early identification and treatment of pneumonia: A call to action. Lancet Glob Health 2016;4: e12–e13. doi: 10.1016/S2214-109X(15)00272-7.
28. Wunderink RG. Guidelines to manage community-acquired pneumonia. Clin Chest Med 2018;39: 723–731. doi: 10.1016/j.ccm.2018.07.006.
29. Xie X, Zhang L, Wu M, Kang Z, Yan H, Zhang X, et al. A retrospective study on the clinical characteristics and radiological features of primary pulmonary lymphoma. Transl Cancer Res 2020;9: 1969–1977. doi: 10.21037/tcr.2020.02.16.
30. Bligh MP, Borgaonkar JN, Burrell SC, MacDonald DA, Manos D. Spectrum of CT findings in thoracic extranodal non-hodgkin lymphoma. Radiographics 2017;37: 439–461. doi: 10.1148/rg.2017160077.
31. Zhang MC, Zhou M, Song Q, Wang S, Shi Q, Wang L, et al. Clinical features and outcomes of pulmonary lymphoma: A single center experience of 180 cases. Lung Cancer 2019;132: 39–44. doi: 10.1016/j.lungcan.2019.04.004.
32. Shen H, Zhou Y. Clinical features and surgical treatment of primary pulmonary lymphoma: A retrospective study. Front Oncol 2022;12: 779395. doi: 10.3389/fonc.2022.779395.
33. Ligero M, Garcia-Ruiz A, Viaplana C, Villacampa G, Raciti MV, Landa J, et al. A CT-based radiomics signature is associated with response to immune checkpoint inhibitors in advanced solid tumors. Radiology 2021;299: 109–119. doi: 10.1148/radiol.2021200928.
34. Wu G, Woodruff HC, Shen J, Refaee T, Sanduleanu S, Ibrahim A, et al. Diagnosis of invasive lung adenocarcinoma based on chest CT radiomic features of part-solid pulmonary nodules: A multicenter study. Radiology 2020;297: E282. doi: 10.1148/radiol.2020209019.
35. Wang H, Wang L, Lee EH, Zheng J, Zhang W, Halabi S, et al. Decoding COVID-19 pneumonia: Comparison of deep learning and radiomics CT image signatures. Eur J Nucl Med Mol Imaging 2021;48: 1478–1486. doi: 10.1007/s00259-020-05075-4.
36. Wang S, Zha Y, Li W, Wu Q, Li X, Niu M, et al. A fully automatic deep learning system for COVID-19 diagnostic and prognostic analysis. Eur Respir J 2020;56: 2000775. doi: 10.1183/13993003.00775-2020.
37. Li L, Wang L, Zeng F, Peng G, Ke Z, Liu H, et al. Development and multicenter validation of a CT-based radiomics signature for predicting severe COVID-19 pneumonia. Eur Radiol 2021;31: 7901–7912. doi: 10.1007/s00330-021-07727-x.
38. Zhang J, Wang G, Ren J, Yang Z, Li D, Cui Y, et al. Multiparametric MRI-based radiomics nomogram for preoperative prediction of lymphovascular invasion and clinical outcomes in patients with breast invasive ductal carcinoma. Eur Radiol 2022;32: 4079–4089. doi: 10.1007/s00330-021-08504-6.
Keywords:

Primary pulmonary lymphoma; Pneumonia; Computed tomography; Radiomics; Differentiation

Supplemental Digital Content

Copyright © 2023 The Chinese Medical Association, produced by Wolters Kluwer, Inc. under the CC-BY-NC-ND license.