18 F-FDG-PET/CT based deep learning model for fully automated prediction of pathological grading for pancreatic ductal adenocarcinoma before surgery

doi:10.21203/rs.3.rs-2578400/v1

Download PDF

Research Article

18 F-FDG-PET/CT based deep learning model for fully automated prediction of pathological grading for pancreatic ductal adenocarcinoma before surgery

https://doi.org/10.21203/rs.3.rs-2578400/v1

This work is licensed under a CC BY 4.0 License

Journal Publication

published 25 May, 2023

Read the published version in EJNMMI Research →

You are reading this latest preprint version

Background ：The determination of pathological grading has a guiding significance for the treatment of pancreatic ductal adenocarcinoma（PDAC）patients. However, there is a lack of an accurate and safe method to obtain pathological grading before surgery. The aim of this study is to develop a deep learning（DL）model based on 18F-FDG-PET/CT for a fully automatic prediction of preoperative pathological grading of pancreatic cancer.

Results ：A total of 370 PDAC patients from January 2016 to September 2021 were collected retrospectively. All patients underwent 18F-FDG-PET/CT examination before surgery and obtained pathological results after surgery. A DL model for pancreatic cancer lesion segmentation was first developed using 100 of these cases and applied to the remaining cases to obtain lesion regions. After that, all patients were divided into training set, validation set and test set according to the ratio of 5:1:1. A predictive model of pancreatic cancer pathological grade was developed using the features computed from the lesion regions obtained by the lesion segmentation model and key clinical characteristics of the patients. Finally, the stability of the model was verified by 7-fold cross-validation.The Dice score of the developed PET/CT based tumour segmentation model for PDAC was 0.89. The area under curve (AUC) of the PET/CT-based DL model developed on the basis of the segmentation model was 0.74, with an accuracy, sensitivity, and specificity of 0.72, 0.73, and 0.72, respectively. After integrating key clinical data, the AUC of the model improved to 0.77, with its accuracy, sensitivity, and specificity boosted to 0.75, 0.77, and 0.73, respectively.

Conclusion ：To the best of our knowledge, this is the first deep learning model to end-to-end predict the pathological grading of PDAC in a fully automatic manner, which is expected to improve clinical decision making.

deep learning

pancreatic cancer

PET/CT

pathological grading

prediction model

Pancreatic cancer is a common malignancy and is the fourth most deadly cancer in the world¹, killing approximately 480,000 people worldwide each year. In the next decade, pancreatic cancer is likely to become the second leading cause of death². Surgery is currently the only treatment that may cure pancreatic cancer. Statistics show³ that 10% of patients with indications for surgical resection have a 5-year survival rate of 24.6% (and 2.9%) for patients who have (and have not) undergone PDAC resection in Stage I.

Predicting pathological grading of pancreatic cancer is an important part of the diagnosis and treatment of pancreatic cancer. Pathological differentiation of pancreatic cancer helps assess the extent, depth, and metastatic status of pancreatic cancer, which is an important basis for determining the best treatment plan and predicting prognosis and plays an important role in guiding surgery and corresponding adjuvant therapy for precise individualized treatment. A study by Golan et al. showed that well-differentiated PDAC was associated with long-term survival after surgery⁴. In contrast, poor differentiation is an independent prognostic factor affecting overall survival⁵. For patients with poorly differentiated PDAC, neoadjuvant therapy may provide longer survival than direct surgery^6-8. The only method currently available to determine PDAC grading preoperatively is ultrasound or CT-guided puncture biopsy. Tumor tissue columns obtained in this manner do not reliably reflect the structural features of the entire lesion due to their high heterogeneity⁹. A study by Larghi et al. showed that the preoperative grading of EUS-FNB had an accuracy of 56%, sensitivity of 41%, and specificity of 78%¹⁰. Therefore, a safe and accurate preoperative method for determining the degree of differentiation of PDAC is needed.

The determination of the pathological grading of PDAC relies on pathological slices as pathological examination is the gold standard for diagnosing the disease. However, since pathological tissue is obtained through invasive puncture or surgery, the pathological results have a pronounced lag. Currently, it is common to diagnose PDAC using medical images, such as CT, MRI, or PET, which can be acquired noninvasively; the diagnosis results are obtained in a shorter time. However, due to low image resolution, the information in medical images is not as clear as that in pathological slices. Also, the variations in equipment and operator lead to unstable imaging results, or at least not as stable as pathological examination. Therefore, how to automatically predict the pathological grading of PDAC through imaging data is a challenging task and has not been investigated in the literature, to the best of our knowledge. In this paper, we attempt to bridge this gap.

Deep learning is a machine learning method that automatically learns features and classifies through the design and use of multi-layer networks^11,12. In recent years, deep learning has been widely applied in medical image analysis, including PET/CT image analysis¹³. PET/CT is a whole-body functional imaging examination, which reflects the malignancy or benignity of lesions through the metabolic activity of cells. Wang et al.¹⁴ studied the use of deep learning models to segment lung cancer in PET/CT images and achieved a high accuracy. Chao et al.¹⁵ used a dual-energy CT based deep learning radiomics model to classify PDAC's lymph node metastasis (LNM) status, and the model's AUC, sensitivity, and specificity were 0.87, 76% and 83%, respectively.

These studies indicate that deep learning models based on PET/CT have a high accuracy and sensitivity. However, current research on pathological grading of PDAC is still limited. Therefore, our goal is to establish a fully automated deep learning model based on ¹⁸F-FDG-PET/CT for predicting preoperative pathological grading of PDAC.

Fig. 1 presents the schematic workflow of the proposed DL model based on PET/CT for pathological grading of patients with PDAC, which consists of multiple processing stages. Below, we elaborate the details related to the workflow, starting with study population, image labeling, to model construction, and finally model testing.

Study population

Patients who underwent pancreatic surgery at the PLA General Hospital from January 2016 to September 2021 and obtained pathological confirmation of PDAC were collected and included in the study according to the inclusion and exclusion criteria, and 370 patients were finally included.

Inclusion criteria: (i) PDAC was pathologically confirmed by radical pancreatic resection; (ii) PDAC was confirmed by pathological biopsy of non-radical pancreatic surgery; (iii) PET/CT of the pancreas was performed within 1 month before surgery. Exclusion criteria: (i) patients had adjuvant treatment such as radiotherapy, chemotherapy and intervention before surgery; (ii) PET/CT images were of poor quality (tumour and borders could not be distinguished with the naked eye or there were artifacts interfering) and could not be used to analyze patients; (iii) other malignant tumours were combined; (iv) there was no significant FDG uptake at the tumour site (SUVmax < 2.5); (v) pathological findings and images could not correspond. Clinical data such as a patient’s age, gender, preoperative CA199 level, tumour location, tumour size (long and short diameter) on PET/CT images, and SUVmax values were also collected.

PET-CT image labeling process

Supplementary Method 1.1 provides detailed information about the PET/CT scanning protocol. The regions of abnormal ¹⁸F-FDG uptake on PET and density abnormality on CT are localized as the lesion region as follows. After the PET/CT image fusion is completed, two experienced PET/CT diagnostic physicians use 3D slicers (version 5.1.0, https://www.slicer.org) software with a threshold of 40% SUVmax to draw out the ROI (Region of Interest) of the target lesion, and all discrepancies are confirmed through discussion. All examination images are interpreted by two senior nuclear medicine specialists (with more than 5 years of PET interpretation experience), including the location of tumor lesion, its relationship with surrounding tissue, the presence of lymph node metastasis, the presence of distant metastasis, and the performance under different sequences.

Constructing the lesion segmentation model

The whole process of building the deep model for lesion segmentation is shown in Figure 1A. 100 cases of annotated PET/CT images of pancreatic cancer were input into the segmentation model for training.

The PET-CT images were first pre-processed: a. Window width and window level (350, 40) were applied to intercept the gray value; b. Each pair of 3D CT series (512*512*H_CT) and 3D PET series (of size 96*96*H_PET, 128*128*H_PET, 168*168*H_PET, 170*170*H_PET) were uniformly resized to 256*256*H_PET; c. For each slice, the gray scale was normalized [0,1]; d. 3 PET slices and 3 CT slices centered around the corresponding location form an 6-channel input to the model.

Model construction: The 6-channel input has a PET part and a CT part, each fed into a 2D-Unet network (Fig.S1.) branch with no shared parameters. The feature vectors of the two 2D-Unets are then concatenated to pass through convolutions, which output the final lesion segmentation mask. The learning rate was set to 1 × 10^-5 and the parameters were updated using the Adam optimiser.

Post-processing the segmentation result: a. A pre-trained model of organ segmentation was added to nnUnet¹⁶ to provide a coarse segmentation of the abdominal organs. The predicted segmentation of pancreatic organ was then fused with the tumour segmentation results to enhance the lesion segmentation performance; b. Medical image analysis techniques including erosion, expansion, and SUVmax 40% threshold segmentation are further applied to obtain the final lesion segmentation results.

Building PDAC pathological grade classification models

Due to the low prevalence of pathological samples with extreme pathological differentiation grades in the clinic, the grades with few samples were merged in this study and all samples were set to two predictive labels: low grade or high grade. Highly, moderate-highly, and moderately differentiated pathologies were defined as low grade; undifferentiated, lowly, and moderate-lowly differentiated were defined as high grade (Fig. S2.). This is similar to the classification method of Wasif and Rochefort et al.^17,18

According to the segmentation result, the lesion regions were cropped out of the 3D data of PET, CT, and segmentation Mask, respectively, and three aligned copies of size 64*64*16 (length*width*height) were obtained. The CT data was intercepted with a window width and window level (350, 40) and normalized to [0,1], and the PET data was normalized to [0,1]. The cropped PET, CT, and Mask were concatenated in the channel dimension to obtain a tensor of size 3*64*64*16 (number_of_channels*length*width*height). The tensor was fed into a Unet3D-based Encoder to extract image feature vectors as shown in Figure 1B.

Cases with clinical data missing ratios greater than 20% were excluded from our study. A total of 21 clinical variables were collected to build predictive models based on clinical experience and literature reports. Subsequently, the individual clinical data was analysed for significance using the Random Forest method (Fig.S3.). Eleven important clinical characteristics including age, BMI, SUVmax, ALT, AST, total bilirubin, direct bilirubin, blood glucose, CEA, CA125 and CA199 were kept. Finally, the clinical data feature vectors were extracted using the MLP through the multilayer perceptron. The part was shown in Figure 1C-D.

Both image features or clinical data features can be used to obtain prediction results for their respective modalities through the fully connected (FC) layer. To obtain better prediction performance, we replaced the last FC layer with a TMC (Trusted Multi-view Classification) ¹⁹ to integrate image features and clinical data features and constructed a PET/CT + Clinical data model. TMC is a new multi-view classification algorithm that dynamically integrates different views at an evidence level to promote classification reliability by considering evidence from each view（Supplementary Method 1.2）. The learning rate was set to 1×10^-5 and the parameters in the feature extractor were updated using the Adam optimiser.

Seven-fold cross validation for model testing

We used a 7-fold cross-validation to better evaluate the generalization ability of the model. This is shown in Figure 2. We divided 370 patients into 7 folds, of which 5 folds were the training set for the model in each training round, 1 fold was the internal validation set and 1 fold was used as the test set to test the final performance of the model. The next round was trained by changing the order of the training, validation and test folds. The final model is obtained by averaging the results of the 7 folds.

Statistical analysis

The clinical data was statistically processed using SPSS 22.0 statistical analysis software: normally distributed measures were expressed as x±s and comparisons between groups were made using the Student-t test. Skewed measures were expressed as median (range) and comparisons of count data were made using the 𝑋^2 test or Fisher's exact probability method.

Dice score was used to evaluate the pancreatic lesion segmentation model. Accuracy, sensitivity, and specificity of the test dataset results were calculated using receiver operator characteristic curve (ROC) for the classification models. P values less than 0.05 were considered statistically significant.

Patient baseline characteristics

From January 2016 to September 2021, 613 consecutive patients with PDAC were retrospectively recruited to our cancer center. Of these, 370 patients (164 women and 206 men; mean age 60.08 ± 9.36 years) were finally screened. These patients were divided into two cohorts based on pathological grading. There were 190 cases in the LG group and 180 cases in the HG group. Table 1 summarizes the baseline characteristics of the patients in the LG and HG groups.

Table.1 Baseline characteristics of patients

Variable	Low grade (n = 190)	High grade (n =180)	P-value
Demographics
Age (y), mean ± SD	59.90±9.22	60.28±9.52	0.698
Gender			0.066
Female	93（49%）	71（39%）
Male	97（51%）	109（61%）
BMI†	23.4（21.47-25.9）	23.44（22.03-25.13）	0.449
Smoke	59（31%）	60（33%）	0.639
Drink alcohol	45（24%）	60（33%）	0.04
Abdominal discomfort	88（46%）	99（55%）	0.095
Weight loss（＞5kg）	39（21%）	57（30%）	0.015
PET/CT report
Tumor size (mm)†
Max	31.5（26.0-45.0）	31（27.0-42.0）	0.728
Min	26（20.0-35.5）	25（19.5-33.0）	0.254
Tumor SUVmax†	5.75（4.4-7.95）	7.5（5.6-9.6）	0.001
Location			0.01
Head-Neck	103（54%）	121（67%）
Body-Tail	87（46%）	59（33%）
Pathological report
Neuroaggression	113（59%）	109（61%）	0.832
Cancerembolus	15（8%）	38（21%）	0.001
Lymph node metastasis	58（31%）	58（32%）	0.834
Laboratory findings†
ALT(U/L)	22.2（11.5-58.7）	17.2（13.2-98.5）	0.159
AST(U/L)	17.70（13.5-49.3）	18.75（13.5-50.8）	0.369
Total-bilirubin(umol/L)	12.4（8.7-18.6）	15.7（8.7-32.7）	0.007
Direct-bilirubin(umol/L)	4.4（2.9-8.8）	4.9（2.7-20.5）	0.037
Glucose (mmol/L)	5.97（5.2-7.0）	5.48（4.8-6.9）	0.162
CEA (μg/L)	2.77（1.8-5.0）	3.02（2.0-4.7）	0.325
CA125(U/mL)	14.46（8.7-25.4）	18.96（11.7-34.7）	0.004
CA199(U/mL)	170.25（47.8-611.2）	229（70.5-738.4）	0.027
† Data in parentheses are the interquartile range
Abbreviation: BMI, Body Mass Index; ALT, alanine aminotransferase; AST, aspartate amino transferase; CEA, Carcinoma Embryonic Antigen; CA 125, carbohydrate antigen 125; CA 199, carbohydrate antigen 199

No significant differences in clinical characteristics were observed between these two cohorts. Alcohol consumption rates were 24.0% (45/190) and 33.3% (60/180) respectively, with a significant difference between the two cohorts (p=0.04). A total of 21.0% (39 out of 190) of the LG group and 30% (87 out of 180) of the HG group had significant weight loss, with a statistically significant difference between the two groups (p=0.015).

All PET/CT reports were independently reviewed and assessed by two experienced PET/CT diagnosticians. The median value of SUVmax was 5.75 (2.2-36.0) in the LG group and 7.5 (3.2-31.1) in the HG group, with a statistically significant difference between the two groups (p=0.001). Tumours in the HG group were more likely to be found in the head and neck of the pancreas than in the LG group 67% vs 54% (p=0.01), a statistically significant difference between the two groups. In laboratory tests, total and direct bilirubin levels were lower in the LG group than in the HG group (12.4 vs 15.7, p=0.007; 4.4 vs 4.9, p=0.037), with a statistically significant difference between the two groups. In terms of tumour marker detection, the LG group had lower levels of CA-125 and CA199 than the HG group (14.46 vs 18.96, p=0.004; 170.25 vs 229, p=0.027).

Performance of lesion segmentation model

The Dice score for the lesion segmentation Unet model in Table.S1. is 0.72. The Dice score for Unet prediction with guidance of organ location (Unet+OL), that is, with the addition of nnUnet-based organ segmentation, was increased to 0.76. The Dice score for Unet+OL prediction with post-processing (Unet+OLP), which is Unet+OL with the addition of post-processing such as erosion, expansion, and threshold segmentation, was improved to 0.89, which is a significant advantage compared to the Unet and Unet+OL models. As can be seen in Figure 3, in the three validated cases, the region of lesions output by Unet+OLP was closer to the Ground Truth (GT) labeling than that of Unet+OL.

Performance of PDAC pathological grade classification model

Regarding testing performance, Figure 4 shows that the AUC of the clinical data model was 0.95 in the training cohort, 0.68 in the validation cohort and 0.68 in the test cohort, while the AUC of the PET/CT model was 0.99 in the training cohort, 0.72 in the validation cohort and 0.74 in the test cohort, which was better than the clinical model. In order to improve the efficacy and accuracy of the model, we combined the clinical model with the PET/CT DL model to build a PET/CT + Clinical data model. The AUC of the PET/CT+ Clinical data model reached 0.99, 0.74, and 0.77 in the training, validation, and test cohorts, respectively.

It can be seen in Table 2 that the accuracy, sensitivity, and specificity of the PET/CT model were 72%, 73%, and 72%, respectively. The accuracy, sensitivity, and specificity of the clinical data model were 66%, 67%, and 66%, respectively. The accuracy, sensitivity, and specificity of PET/CT+ Clinical data model were 75%, 77%, and 73% respectively. The model that integrates both image and clinical data achieved the best performance.

Table.2 The performance comparison of different models

Models	Cohorts	AUC	ACC	SENS	SPEC	PPV	NPV
PET/CT	Train	0.99	0.97	0.97	0.97	0.97	0.96
	Val	0.72	0.73	0.72	0.74	0.75	0.70
	Test	0.74	0.72	0.73	0.72	0.72	0.72
Clinical	Train	0.95	0.91	0.90	0.91	0.92	0.89
	Val	0.68	0.67	0.66	0.69	0.75	0.59
	Test	0.68	0.66	0.67	0.66	0.69	0.63
PET/CT+Clinical	Train	0.99	0.98	0.98	0.97	0.98	0.98
	Val	0.73	0.76	0.75	0.76	0.78	0.73
	Test	0.77	0.75	0.77	0.73	0.73	0.76
AUC, area under receiver operating characteristic curve; ACC, accuracy; SENS, sensitivity; SPEC, specificity; PPV, positive predictive value; NPV, negative predictive value

Currently, research on deep learning models for PDAC mainly focuses on the disease's differential diagnosis, preoperative staging, and prognostic analysis. Wei et al.²⁰ used a combination of machine learning and deep learning algorithms to extract features from PET/CT images to predict the difference between PDAC and autoimmune pancreatitis, developing a multi-domain fusion model with an overall performance of AUC, accuracy, sensitivity, and specificity of 0.96, 0.90, 0.88, and 0.93, respectively. Bian et al.²¹ developed and validated an automated preoperative AI algorithm for tumor and lymph node segmentation in CT imaging to predict LN metastasis in PDAC patients. Lee et al.²² developed a deep learning model based on clinical data to predict postoperative survival in pancreatic cancer patients, and the model's performance in predicting 2-year OS was comparable to AJCC (AUC, 0.67; P=0.35), and it was better than AJCC in predicting 1-year recurrence free survival (AUC, 0.54; P=0.049). Yao et al.²³ used deep learning to analyze preoperative multi-phase CT and developed imaging-derived biomarkers for predicting overall survival in PDAC, which can be used to predict OS in resectable PDAC patients.

Research has found that the pathological grade of PDAC is largely determined by the fibrous matrix quality in its stroma. Tumors with lower differentiation have more fibrous matrix and occupy more of the contrast agent²⁴. This provides a principle for pathological grading of PDAC through imaging studies. For example, Tikhonova et al.²⁵ used a machine learning algorithm to establish a diagnostic model for image-based PDAC grading based on preoperative CT, using data from 91 patients to establish a diagnostic model. The AUC for pathological grading ≥ 2 (or 3) was 0.75 (or 0.66). Na et al.²⁴ developed and validated a radiological feature based on contrast-enhanced computed tomography for preoperative prediction of histological grading of PDAC, and the AUC of the final validation set was 0.770.

The above studies are all based on CT. PET/CT, integrating PET and CT into the same device and program, characterizes the lesion from different aspects and provides metabolic information from the former and detailed anatomical information from the latter, which makes the PET/CT image to have both good clarity and a strong ability to distinguish between lesion tissue and normal tissue²⁶. Some research has shown that, based on 102 patients with histologically confirmed PDAC, FDG uptake is related to the invasiveness of pancreatic cancer, and SUVmax is significantly related to pathological grading²⁷. Therefore, our goal is to unleash the potential of PET/CT in pathological grading of PDAC.

In this study, in order to achieve the prediction of fully automatic PDAC pathological grading and reduce the impact of confounding factors in PET/CT images as much as possible, we first developed a deep learning model for PDAC lesion segmentation. Due to the presence of FDG uptake in organs such as the liver, adrenal gland, small intestine, and bladder in addition to the pancreatic lesion, the model showed abnormal segmentation in the initial training (Figure S4). In order to increase the accuracy of segmentation, nnUnet's organ segmentation pre-training model was added to provide a rough segmentation of abdominal organs, which are filtered out to enhance the performance of pancreatic lesion segmentation. While nnUnet¹⁶ is an excellent image segmentation model that demonstrated good performance in the field of medical image segmentation especially with a large amount of training images, it is difficult to directly obtain a good segmentation model due to the small number of pancreatic cancer sections in this study. We first trained the nnUnet model of pancreatic organ segmentation using abdominal CT sections of healthy people, preserved the parameters of the model, and applied it to the target cases. The presence of tumors greatly reduced the segmentation results of the nnUnet model, but the preliminary localization of the patient's pancreas can still be obtained. The addition of nnUnet as a filter increased the Dice score of the model from 0.72 to 0.76. In the post-processing, we added corrosion, expansion, SUVmax 40% threshold segmentation and other post-operations, and finally increased the Dice score to 0.86 (Table.S1.). Through these steps, the segmentation model's performance was able to achieve an acceptable level.

In building a classification model for PET/CT images, to incorporate more effective information while minimizing the effect of segmentation errors, we cropped out a 3D patch from the raw image centered around the segmented mask. The accuracy of the final PET/CT-based classification model was 72%, sensitivity was 73%, and specificity was 72%.

To further boost the classification performance, we resorted to clinical indicators that are important for revealing pancreatic cancer characteristics. For example, CA199 is closely related to the prognosis of pancreatic cancer patients^28,29. A study found that CA199 produced by pancreatic cancer cell lines in vitro is associated with histological differentiation in nude mice in vivo³⁰. Therefore, we extracted key clinical data indicators such as CA199 and combined them with the PET/CT model to optimize the prediction accuracy. In our early experiment, we used fully connected layers to connect PET/CT with the extracted clinical features, and achieved a minor performance improvement; in some folds in cross validation, the combined model performed worser than the original PET/CT model. Therefore, in order to better integrate the features of PET/CT and clinical data, we used TMC¹⁹ to improve the reliability of classification. This model parameterizes different data and combines them based on Dempster-Shafer theory, which improves the reliability and robustness of the classification model and improves the performance of the model (refer to Supplementary Method 1.2). Finally, to test the generalization ability of the model, we utilized 7-fold cross-validation. The final PET/CT + Clinical data model achieved an accuracy of 75%, sensitivity of 77%, and specificity of 73%. The deep learning model of PET/CT + Clinical data has a significant improvement compared to the traditional EUS-FNB¹⁰ with an accuracy of 56% and sensitivity of 41%.

Our research has a few limitations. Firstly, our data is from a single center and lacks external datasets for validation and evaluation. However, because the medical images were acquired using three different PET machines, it increases the generalizability. Additionally, using 7-fold cross-validation ensured the stability of the final model, which partially compensated for the lack of external validation sets. Secondly, all the data included in this study were from surgical patients, so the pathological differentiation was concentrated; however, the data for patients with excessive malignancy was relatively scarce, which did not make the model to show more significant discriminability. Finally, this study only explored the relationship between imaging features and pathological differentiation of PDAC but did not investigate the survival outcome of the patients, which is more concerned by patients and surgeons. Further research is needed to study the survival of patients.

Our developed deep learning model can automatically analyze PET/CT images without human intervention, reducing subjective errors and improving the accuracy and reliability of grading. It is important to note that deep learning models cannot completely replace professional knowledge of imaging and pathology. Before using deep learning models to predict pathological grading, professional knowledge needs to be combined for interpretation and evaluation to ensure the accuracy and reliability of the results. In conclusion, deep learning-based PET/CT has a great potential in pathological grading of pancreatic cancer, and more clinical studies are needed to prove its safety and effectiveness before deep learning models can replace traditional pathological staging methods.

To the best our knowledge, this is the first report of using a DL model for preoperative prediction of PDAC pathological grading using PET/CT. The model's predictive performance was improved by combining features of PET/CT and key clinical data.

PDAC	Pancreatic ductal adenocarcinoma
DL	Deep learning
AUC	Area under curve
SUV	Standardized uptake value
LNM	Lymph node metastasis
LG	Low grade
HG	High grade
FC	Fully connected
TMC	Trusted Multi-view Classification
ROC	Receiver operator characteristic curve
GT	Ground truth
OL	Organ location
OLP	Organ location with post-processing
ACC	Accuracy
SENS	Sensitivity
SPEC	Specificity
PPV	Positive predictive value
NPV	Negative predictive value

Supplementary Information

Additional file 1: Fig.S1. Unet image segmentation network. Fig.S2. Distribution data of pathological differentiation degree in medical records. Fig.S3. Random forest analysis. Fig.S4. Examples of segmentation model before and after adding nnUet. Table.S1. The performance comparison of segmentation process.

Acknowledgements

Not applicable.

Author Contributions

All authors were involved in the study conception and design. GZ, YZL, ZZW, LD, FW, BXX were involved in acquisition of data. GZ, CKB, YZ, S. KZ participated in the development and testing of deep learning models. GZ, CKB, ZZW, YZ were involved in analysis and interpretation of the data. GZ, CKB and S. KZ were involved in drafting of the manuscript. RL is responsible for review, editing and supervision. All authors read and approved the final manuscript.

Funding

This work was supported by the National Key R&D Program of China (2021ZD0113301)；

Availability of data and materials

The relevant images and clinical data from this study are not available because they contain private patient information. However, such data can be obtained through agency approval and signed data use agreements and/or signed material transfer agreements.

Ethics approval and consent to participate

Approved by the Ethics Review Committee of the local institution of the PLA General Hospital, and the requirement of written informed consent was waived. All procedures involving human participants in this study comply with institutional and/or national Research Council ethical standards as well as the 1964 Declaration of Helsinki and its later amendments or similar ethical standards.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.

Siegel RL, Miller KD, Jemal A. Cancer statistics, 2018. CA: a cancer journal for clinicians. 2018;68(1):7-30.
Hartwig W, Werner J, Jäger D, Debus J, Büchler MW. Improvement of surgical results for pancreatic cancer. The Lancet Oncology. 2013;14(11):e476-e485.
Strobel O, Neoptolemos J, Jäger D, Büchler MW. Optimizing the outcomes of pancreatic cancer surgery. Nature reviews Clinical oncology. 2019;16(1):11-26.
Golan T, Sella T, Margalit O, et al. Short- and Long-Term Survival in Metastatic Pancreatic Adenocarcinoma, 1993-2013. Journal of the National Comprehensive Cancer Network : JNCCN. 2017;15(8):1022-1027.
Han SH, Heo JS, Choi SH, et al. Actual long-term outcome of T1 and T2 pancreatic ductal adenocarcinoma after surgical resection. International journal of surgery (London, England). 2017;40:68-72.
de Geus SW, Eskander MF, Bliss LA, et al. Neoadjuvant therapy versus upfront surgery for resected pancreatic adenocarcinoma: A nationwide propensity score matched analysis. Surgery. 2017;161(3):592-601.
Shubert CR, Bergquist JR, Groeschl RT, et al. Overall survival is increased among stage III pancreatic adenocarcinoma patients receiving neoadjuvant chemotherapy compared to surgery first and adjuvant chemotherapy: An intention to treat analysis of the National Cancer Database. Surgery. 2016;160(4):1080-1096.
Versteijne E, Vogel JA, Besselink MG, et al. Meta-analysis comparing upfront surgery with neoadjuvant treatment in patients with resectable or borderline resectable pancreatic cancer. The British journal of surgery. 2018;105(8):946-958.
Gerlinger M, Rowan AJ, Horswell S, et al. Intratumor heterogeneity and branched evolution revealed by multiregion sequencing. The New England journal of medicine. 2012;366(10):883-892.
Larghi A, Correale L, Ricci R, et al. Interobserver agreement and accuracy of preoperative endoscopic ultrasound-guided biopsy for histological grading of pancreatic cancer. Endoscopy. 2015;47(4):308-314.
Elemento O, Leslie C, Lundin J, Tourassi G. Artificial intelligence in cancer research, diagnosis and therapy. Nature reviews Cancer. 2021;21(12):747-752.
Kleppe A, Skrede OJ, De Raedt S, Liestøl K, Kerr DJ, Danielsen HE. Designing deep learning studies in cancer diagnostics. Nature reviews Cancer. 2021;21(3):199-211.
Choi H. Deep Learning in Nuclear Medicine and Molecular Imaging: Current Perspectives and Future Directions. Nuclear medicine and molecular imaging. 2018;52(2):109-118.
Wang S, Mahon R, Weiss E, et al. Automated Lung Cancer Segmentation Using a PET and CT Dual-Modality Deep Learning Neural Network. International journal of radiation oncology, biology, physics. 2023;115(2):529-539.
An C, Li D, Li S, et al. Deep learning radiomics of dual-energy computed tomography for predicting lymph node metastases of pancreatic ductal adenocarcinoma. European journal of nuclear medicine and molecular imaging. 2022;49(4):1187-1199.
Isensee F, Petersen J, Klein A, et al. nnU-Net: Self-adapting Framework for U-Net-Based Medical Image Segmentation. 2018.
Wasif N, Ko CY, Farrell J, et al. Impact of tumor grade on prognosis in pancreatic cancer: should we include grade in AJCC staging? Annals of surgical oncology. 2010;17(9):2312-2320.
Rochefort MM, Ankeny JS, Kadera BE, et al. Impact of tumor grade on pancreatic cancer prognosis: validation of a novel TNMG staging system. Annals of surgical oncology. 2013;20(13):4322-4329.
Han Z, Zhang C, Fu H, Zhou JT. Trusted Multi-View Classification with Dynamic Evidential Fusion. IEEE transactions on pattern analysis and machine intelligence. 2022;Pp.
Wei W, Jia G, Wu Z, et al. A multidomain fusion model of radiomics and deep learning to discriminate between PDAC and AIP based on (18)F-FDG PET/CT images. Japanese journal of radiology. 2022:1-11.
Bian Y, Zheng Z, Fang X, et al. Artificial Intelligence to Predict Lymph Node Metastasis at CT in Pancreatic Ductal Adenocarcinoma. Radiology. 2023;306(1):160-169.
Lee W, Park HJ, Lee HJ, et al. Preoperative data-based deep learning model for predicting postoperative survival in pancreatic cancer patients. International journal of surgery (London, England). 2022;105:106851.
Yao J, Cao K, Hou Y, et al. Deep Learning for Fully Automated Prediction of Overall Survival in Patients Undergoing Resection for Pancreatic Cancer: A Retrospective Multicenter Study. Annals of surgery. 2022.
Chang N, Cui L, Luo Y, Chang Z, Yu B, Liu Z. Development and multicenter validation of a CT-based radiomics signature for discriminating histological grades of pancreatic ductal adenocarcinoma. Quantitative imaging in medicine and surgery. 2020;10(3):692-702.
Tikhonova VS, Karmazanovsky GG, Kondratyev EV, et al. Radiomics model-based algorithm for preoperative prediction of pancreatic ductal adenocarcinoma grade. European radiology. 2022.
Beyer T, Antoch G, Müller S, et al. Acquisition protocol considerations for combined PET/CT imaging. Journal of nuclear medicine : official publication, Society of Nuclear Medicine. 2004;45 Suppl 1:25s-35s.
Wang Z, Chen JQ, Liu JL, Qin XG, Huang Y. FDG-PET in diagnosis, staging and prognosis of pancreatic carcinoma: a meta-analysis. World journal of gastroenterology. 2013;19(29):4808-4817.
Williams JL, Kadera BE, Nguyen AH, et al. CA19-9 Normalization During Pre-operative Treatment Predicts Longer Survival for Patients with Locally Progressed Pancreatic Cancer. Journal of gastrointestinal surgery : official journal of the Society for Surgery of the Alimentary Tract. 2016;20(7):1331-1342.
Chen Y, Shao Z, Chen W, et al. A varying-coefficient cox model for the effect of CA19-9 kinetics on overall survival in patients with advanced pancreatic cancer. Oncotarget. 2017;8(18):29925-29934.
Iwamura T, Taniguchi S, Kitamura N, et al. Correlation between CA19-9 production in vitro and histological grades of differentiation in vivo in clones isolated from a human pancreatic cancer cell line (SUIT-2). Journal of gastroenterology and hepatology. 1992;7(5):512-519.

Additionalfile1.docx

Download PDF

Journal Publication

published 25 May, 2023

Read the published version in EJNMMI Research →

Reviewers agreed at journal
17 Feb, 2023
Editor assigned by journal
14 Feb, 2023
First submitted to journal
13 Feb, 2023

You are reading this latest preprint version

18 F-FDG-PET/CT based deep learning model for fully automated prediction of pathological grading for pancreatic ductal adenocarcinoma before surgery

Status:

Journal Publication

Version 1

Abstract

Figures

Background

Materials And Methods

Study population

PET-CT image labeling process

Constructing the lesion segmentation model

Building PDAC pathological grade classification models

Seven-fold cross validation for model testing

Statistical analysis

Results

Patient baseline characteristics

Performance of lesion segmentation model

Performance of PDAC pathological grade classification model

Discussion

Conclusions

Abbreviations

Declarations

References

Supplementary Files

Status:

Journal Publication

Version 1