Professional Documents
Culture Documents
s11547-023-01625-6
s11547-023-01625-6
s11547-023-01625-6
https://doi.org/10.1007/s11547-023-01625-6
ABDOMINAL RADIOLOGY
Received: 14 November 2022 / Accepted: 27 March 2023 / Published online: 28 April 2023
© Italian Society of Medical Radiology 2023
Abstract
Background Accurate preoperative clinical staging of gastric cancer helps determine therapeutic strategies. However, no
multi-category grading models for gastric cancer have been established. This study aimed to develop multi-modal (CT/EHRs)
artificial intelligence (AI) models for predicting tumor stages and optimal treatment indication based on preoperative CT
images and electronic health records (EHRs) in patients with gastric cancer.
Methods This retrospective study enrolled 602 patients with a pathological diagnosis of gastric cancer from Nanfang
hospital retrospectively and divided them into training (n = 452) and validation sets (n = 150). A total of 1326 features were
extracted of which 1316 radiomic features were extracted from the 3D CT images and 10 clinical parameters were obtained
from electronic health records (EHRs). Four multi-layer perceptrons (MLPs) whose input was the combination of radiomic
features and clinical parameters were automatically learned with the neural architecture search (NAS) strategy.
Results Two two-layer MLPs identified by NAS approach were employed to predict the stage of the tumor showed greater
discrimination with the average ACC value of 0.646 for five T stages, 0.838 for four N stages than traditional methods with
ACC of 0.543 (P value = 0.034) and 0.468 (P value = 0.021), respectively. Furthermore, our models reported high prediction
accuracy for the indication of endoscopic resection and the preoperative neoadjuvant chemotherapy with the AUC value of
0.771 and 0.661, respectively.
Conclusions Our multi-modal (CT/EHRs) artificial intelligence models generated with the NAS approach have high
accuracy for tumor stage prediction and optimal treatment regimen and timing, which could facilitate radiologists and
gastroenterologists to improve diagnosis and treatment efficiency.
Abbreviations
EGC Early gastric cancer
EHRs Electronic health records
Shangqing Liu and Weiqi Liang share co-first authorship.
LAGC Locally advanced gastric cancer
* Tao Chen AUC Area under the receiver operating characteristic
drchentao@163.com curve
1 CA125 Cancer antigen 125
School of Biomedical Engineering, Southern Medical
University, Guangzhou 510515, China CA19-9 Carbohydrate antigen 19-9
2 CEA Carcinoembryonic antigen
Department of General Surgery and Guangdong Provincial
Key Laboratory of Precision Medicine for Gastrointestinal CNN Convolutional Neural Network
Tumor, Nanfang Hospital, The First School of Clinical DL Deep learning
Medicine, Southern Medical University, Guangzhou 510515, MLP Multi-layer perceptron
Guangdong, China NAS Neural architecture search
3
Department of Radiology, Nanfang Hospital, Southern
Medical University, Guangzhou 510515, China
4
Department of Gastrointestinal and Hernia Surgery, Ganzhou
Hospital‑Nanfang Hospital, Southern Medical University,
Ganzhou 341000, Jiangxi, China
13
Vol.:(0123456789)
510 La radiologia medica (2023) 128:509–519
13
La radiologia medica (2023) 128:509–519 511
13
512 La radiologia medica (2023) 128:509–519
Fig. 2 The pipeline of model development and model application. connected neural network was developed using the neural architecture
A Collection of clinical data. Digital data of patients were extracted search strategy. D Model evaluation. The classification abilities
from electronic health records (EHRs) and three-dimensional (3D) of our models were assessed by the ROC and classification metrics
segmentation of tumors in CT images were performed. B Radiomic such as accuracy, sensitivity and specificity. During the model
feature extraction and clinical characteristics preprocessing. application process, the new patient information was processed by
Radiomics features, including intensity, shape, and texture features, the constructed model and got the patient-level tumor stage and the
were extracted from ROIs. C Model construction. Radiomic indication of optimal treatment decisions
features and clinical characteristics fused. Finally, a two-layer fully
In total, we extracted 1316 features from each patient. the pipeline of image processing, model development,
Features were divided into three groups: (1) First-order and model application. We use NAS approach to develop
statistics (n = 252); (2) shape (n = 14); (3) texture based four MLP (multi-layer perceptron) models for four tasks:
(n = 1050), including GLCM (n = 336), GLSZM (n = 224), (1) prediction of T-stage; (2) prediction of N-stage; (3)
GLRLM (n = 224), GLDM (n = 196) and NGTDM (n = 70). prediction of the indication for endoscopic procedure; (4)
prediction of the indication for neoadjuvant chemotherapy.
Clinical features extraction We applied NAS using the python package of Optuna, an
open-source hyperparameter optimization framework to
Clinical information collected included age, gender, tumor automate hyperparameter search [19]. Details of model
location, T stage, N stage, degree of differentiation, and development and settings are described in Appendix E1.
tumor markers such as CEA, CA19-9, CA72-4, CA12-5
from EHRs. In total, we extracted 10 clinical features from Model evaluation
each patient. All the values were categorized using integers.
The area under the receiver operating characteristic curve
Model algorithms and development (AUC) was used as the performance measure. Accuracy,
sensitivity, and specificity were calculated too. We
After extracting the clinical and visual features from each calculated their macro average for evaluation. The true
patient, we concatenated these two features and got the positive definition of a specific category is accurately
final multi-modal input which could make the model learn predicted as the category, and the true negative definition is
effectively with more diverse features. Figure 2 presents accurately predicted as one of the remaining four categories.
13
La radiologia medica (2023) 128:509–519 513
Youden Index was calculated to determine the best cutoff Table 1 Patient characteristics in the training and validation cohorts
value to predict the outcomes. Training cohort Validation cohort P value
(N = 452) (N = 150)
Statistical analysis
Age (year) 56.0 ± 11.5 54.5 ± 10.8 0.155
Gender
Statistical analyses were performed using python software
Male 298 (65.9%) 105 (70.0%) 0.413
(version 3.8.3, https://w
ww.p ython.o rg/) with Scipy package
Female 154 (34.1%) 45 (30.0%)
(version 1.5.1, https://www.scipy.org/) or SPSS software
Tumor location
(version 24.0; SPSS Inc., Chicago, IL). Differences between
Up 93 (20.6%) 30 (20.0%) 0.324
cohorts were assessed using t-tests or Mann–Whitney U
Medium 93 (20.6%) 23 (15.3%)
tests for continuous variables; chi-squared tests or Fisher’s
Low 266 (58.8%) 97 (64.7%)
exact test were applied for categorical variables. Receiver
Tumor grade
operating characteristic (ROC) curves were used to display
High 22 (4.9%) 7 (4.7%) 0.78
and evaluate model performance.
Medium 75 (16.6%) 30 (20.0%)
Low 350 (77.4%) 112 (74.7%)
Undifferentiated 5 (1.1%) 1 (0.7%)
Results
cT stage
T1a 16 (3.5%) 3 (2.0%) 0.776
Baseline information of participants
T1b 39 (8.6%) 13 (8.7%)
T2 57 (12.6%) 15 (10.0%)
The 602 patient cohort from the center was divided into a
T3 81 (17.9%) 26 (17.3%)
training cohort (N = 452) and a validation cohort (N = 150).
T4a 245 (54.2%) 86 (57.3%)
Detailed clinicopathological baseline data and comparison
T4b 14 (3.1%) 7 (4.7%)
results are shown in Table 1. The training cohort consisted of
cN stage
298 males and 154 females (mean age 56.0 ± 11.5 years, age
N0 233 (51.5%) 76 (50.7%) 0.656
range 22–83 years), and the validation cohort consisted of
N1 106 (23.5%) 33 (22.0%)
105 males and 45 females (mean age 54.5 ± 10.8 years, age
N2 82 (18.1%) 26 (17.3%)
range 22–79 years). There were no significant differences
N3 31 (6.9%) 15 (10.0%)
in tumor location (longitudinal), tumor differentiation,
cM stage
preoperative CA 19–9 level, and CA 125 level, and
M0 441 (97.6%) 148 (98.7%) 0.632
significant differences in preoperative CEA level and CA
M1 11 (2.4%) 2 (1.3%)
72–4 between these two cohorts.
cTNM
I 94 (20.8%) 23 (15.3%) 0.059
Prediction of tumor stage and feature importance
II A 10 (2.2%) 7 (4.7%)
analysis
II B 125 (27.7%) 34 (22.7%)
III 195 (43.1%) 72 (48.0%)
Based on the Yonden Index, we determined the model
IV A 17 (3.8%) 12 (8.0%)
threshold for each prediction task. The overall accuracy for
IV B 11 (2.4%) 2 (1.3%)
predicting T-stage in the training and validation cohort were
pT stage
0.803 and 0.646, respectively, and the macro-average AUC
T1a 33 (7.3%) 11 (7.3%) 0.912
values were 0.892 and 0.716 (Table E2-3, Fig. 3A, B). In
T1b 51 (11.3%) 13 (8.7%)
the prediction model of lymph node metastasis, the overall
T2 53 (11.7%) 15 (10.0%)
accuracy in the training and validation cohort were 0.669
T3 43 (9.5%) 17 (11.3%)
and 0.838. respectively, and the macro-average AUC values
T4a 234 (51.8%) 81 (54.0%)
were 0.712 and 0.698 (Table E4-5, Fig. 3C, D).
T4b 38 (8.4%) 13 (8.7%)
To better understand the landscape of clinical parameters
pN stage
and radiomic features in the prediction model, we took
N0 184 (40.7%) 49 (32.7%) 0.294
the CNN model as a dark box and got the weight of each
N1 68 (15.0%) 27 (18.0%)
input by permutation technique. In the prediction task of
N2 81 (17.9%) 27 (18.0%)
T-stage, original image derived dependence variance (DV)
N3a 67 (14.8%) 31 (20.7%)
of GLDM applying LoG filter with Gaussian kernel size
N3b 52 (11.5%) 16 (10.7%)
of 4 (log-4_gldm_DependenceVariance) weighted the most
important in the forecasting system (Fig. 4A). The top three
13
514 La radiologia medica (2023) 128:509–519
13
La radiologia medica (2023) 128:509–519 515
Fig. 3 ROC curves of the prediction model. A infiltration depth model in the training cohort; D lymph node metastasis prediction
prediction model in the training cohort; B infiltration depth prediction model in the validation cohort
model in the validation cohort; C lymph node metastasis prediction
on the surgeon's accurate preoperative clinical staging information to classify T-stage in binary categories for GC
and risk stratification. Moreover, early identification of such as T2 vs T3/T4 and T1/T2/T3 vs T4 successfully with
LAGC patients before surgery can improve the response the AUC values of 0.818–0.825 and 0.813–0.840 in the
rate of neoadjuvant chemotherapy and provide further testing cohort [11, 22]; Chen et al. and Wang et al. applied
treatment for patients who lost the chance of surgery. the DL method to diagnostic analysis such as recurrence
According to the literature, the accuracy rate of Radiologists prediction and metastasis identification [23, 24]. However,
subjectively determining the depth of invasion and lymph most studies adopted a single examination method, and did
node metastasis of GC based on CT images are about 70%, not integrate multi-modal examination methods.
which is unsatisfactory [20, 21]. In previous studies, Wang In this study, we combined enhanced CT visual data with
et al. and Meng et al. used traditional radiomics and clinical the EHRs data to construct multi-modal models to focus on
13
516 La radiologia medica (2023) 128:509–519
Fig. 4 Top 10 important features for gastric cancer A infiltration depth prediction model; B lymph node metastasis prediction model
Fig. 5 Spearman’s correlation plots of tumor infiltration depth (A) and lymph node metastasis (B) with their top important feature, respectively
the prediction of specific stages of GC into five T categories growth, which from the view of graphic processing is
and four N categories as well as the identification to recive the relationship between a voxel and neighboring voxels.
endoscopic resection and neoadjuvant chemotherapy Consequently, the histogram-like function of GLDM
preoperatively to simulate real-world clinical decision- were able to illustrate the color transitions between
making. Given a limited number of training samples, adjacent pixels with better performance. As for clinical
training multiple sophisticated deep learning models are parameters, CA19-9 gives a robust contribution to the
very challenging due to the choice of the hyperparameters classification task. Carbohydrate antigen 19-9 has been
and construction of the networks architecture. Fortunately, widely used for assisting diagnosis and treatment efficacy
we can use the NAS approach adaptively to design the best evaluation in digestive tract malignancies like pancreatic
architecture and hyperparameters to tackle these challenges. cancer or cholangiocarcinoma [25, 26]. The association
From the feature importance analysis, we can see that of elevated CA19-9 levels with gastric carcinoma has
the top 10 important features have both radiomic features been presented in a case report and other studies [27,
from CT images and clinical parameters from EHRs. This 28]. Elevated CA19-9 levels have been significantly
means that the multi-modal (CT/EHRs) information is correlated with lymph node metastasis, vascular invasion,
more helpful for gastric cancer diagnosis and treatment and liver metastasis. Our results indicated that CA19-9
decisions compared with single modal CT images or aids in T-stage and N-stage differentiation, consistent with
EHRs. In terms of visual information, radiomic features previous researches [29, 30].
extracted from the original image filtered by wavelet are We found that the model performed poorly in
less critical than LoG and the original image. The most intermediate classes such as T3 or T4a (Table E2–E3). An
significant feature in this study originated from the GLDM explanation for this phenomenon was that the category of
method of measuring variance independence size in CT T-stage generally depends on the location and volume of
images. We considered this finding to be relevant to the the primary tumor lesion. The difference between the sizes
interpretation of images. T-stage describes the expansion of some intermediate tumors may be subtle. Secondly, there
of the lesion border within the gastric wall during tumor existed a severe class imbalance in our data, consistent with
13
La radiologia medica (2023) 128:509–519 517
Fig. 6 ROC curves of the optimal clinical decision model and Top 10 important features of each model, respectively. A ROC curve of the IER
model; B ROC curve of the INC model; C important features in the IER model; D important features in the INC model
the real-world clinical practice in our country that the GC need to be considered. Firstly, this was a retrospective study
patients graded T1/T2 is relatively fewer than T3/T4 [31]. conducted at a single center with inter-class imbalance.
The key strengths of this study include the use of multi- Secondly, there is still room for the accuracy and overall
modal (CT/EHRs) information with the combination of performance of our model to improve. Thirdly, the prediction
visual features and clinical features, and the use of NAS of the indications of neoadjuvant and endoscopic therapy were
approach to develop p the artificial intelligence model which only a stimulation of clinicians in clinical decision-making
alleviates the challenge for the choice of hyperparameters and based on clinical tumor stage prior to surgery, thus it can only
networks architecture during training. In detail, we built two provide reference rather than medical instructions. Fourthly,
multi-category prediction model which is nearly consistent a multicenter study with a larger dataset is needed to further
with pathological findings using radiomics features from CT validate our models’ reproducibility and generalization ability
images and clinical parameters. This will help doctors save a in future studies.
certain amount of time spend on diagnosis, so as to gain more
treatment time for patients. To our knowledge, our method is
the first work combining the radiomics and NAS approach to
develop the AI models, which alleviates the challenge for the
choice of hyperparameters and network architecture during
training. However, there are also some potential limitations
13
518 La radiologia medica (2023) 128:509–519
13
La radiologia medica (2023) 128:509–519 519
surgical resection of local primary gastrointestinal stromal tumors early detection and monitoring progression of gastric cancer: a
based on deep learning. EBioMedicine 39:272–279 multiphase study. JAMA Surg 155(7):572–579
24. Wang X, Chen Y, Gao Y et al (2021) Predicting gastric cancer 30. Song XH, Liu K, Yang SJ et al (2020) Prognostic value of changes
outcome from resected lymph node histopathology images using in preoperative and postoperative serum CA19-9 levels in gastric
deep learning. Nat Commun 12(1):1637 cancer. Front Oncol 10:1432
25. Carpelan-Holmström M, Louhimo J, Stenman UH, Alfthan H, 31. Tsendsuren T, Jun SM, Mian XH (2006) Usefulness of endoscopic
Haglund C (2002) CEA, CA 19-9 and CA 72-4 improve the ultrasonography in preoperative TNM staging of gastric cancer.
diagnostic accuracy in gastrointestinal cancers. Anticancer Res World J Gastroenterol 12(1):43–47
22(4):2311–2316
26. Ballehaninna UK, Chamberlain RS (2012) The clinical utility Publisher's Note Springer Nature remains neutral with regard to
of serum CA 19-9 in the diagnosis, prognosis and management jurisdictional claims in published maps and institutional affiliations.
of pancreatic adenocarcinoma: an evidence based appraisal. J
Gastrointest Oncol 3(2):105–119 Springer Nature or its licensor (e.g. a society or other partner) holds
27. Feng F, Tian Y, Xu G et al (2017) Diagnostic and prognostic value exclusive rights to this article under a publishing agreement with the
of CEA, CA19-9, AFP and CA125 for early gastric cancer. BMC author(s) or other rightsholder(s); author self-archiving of the accepted
Cancer 17(1):737 manuscript version of this article is solely governed by the terms of
28. Jo JC, Ryu MH, Koo DH et al (2013) Serum CA 19–9 as a such publishing agreement and applicable law.
prognostic factor in patients with metastatic gastric cancer. Asia
Pac J Clin Oncol 9(4):324–330
29. Guo X, Lv X, Ru Y et al (2020) Circulating exosomal gastric
cancer-associated long noncoding RNA1 as a biomarker for
13