The OCDA-Net: a 3D convolutional neural network-based system for classification and staging of ovarian cancer patients using FDG PET/CT examinations

doi:10.21203/rs.3.rs-3099695/v1

Download PDF

Research Article

The OCDA-Net: a 3D convolutional neural network-based system for classification and staging of ovarian cancer patients using FDG PET/CT examinations

https://doi.org/10.21203/rs.3.rs-3099695/v1

This work is licensed under a CC BY 4.0 License

Journal Publication

published 28 Sep, 2023

Read the published version in Annals of Nuclear Medicine →

You are reading this latest preprint version

Objective

To create the 3D convolutional neural network (CNN)-based system that can use whole-body FDG PET for recurrence/post-therapy surveillance in ovarian cancer (OC).

Methods

This study 1224 image sets from OC patients who underwent whole-body FDG PET/CT at Kowsar hospital between April 2019 and May 2022 were investigated. For recurrence/post-therapy surveillance, diagnostic classification as cancerous, and non-cancerous and staging as stage III, and stage IV were determined by pathological diagnosis and specialists’ interpretation. New deep neural network algorithms, the OCDAc-Net, and the OCDAs-Net were developed for diagnostic classification and staging of OC patients using PET/CT images. Examinations were divided into independent training (75%), validation (10%), and testing (15%) subsets.

Results

This study included 37 women (mean age, 56.3 years; age range, 36–83 years). Data augmentation techniques were applied to the images in two phases. There were 1224 image sets for diagnostic classification and staging. For the test set, 170 image sets were considered for diagnostic classification and staging. The OCDAc-Net areas under the receiver operating characteristic curve (AUCs) and overall accuracy for diagnostic classification were 0.990 and 0.92, respectively. The OCDAs-Net achieved areas under the receiver operating characteristic curve (AUCs) of 0.995 and overall accuracy of 0.94 for staging.

Conclusions

The proposed 3D CNN-based models provide potential tools for recurrence/post-therapy surveillance in OC. The OCDAc-Net and the OCDAs-Net model provide a new prognostic analysis method that can utilize PET images without pathological findings for diagnostic classification and staging.

3D Convolutional Neural Network

Classification

Staging

Ovarian Cancer

PET/CT

The primary gynecologic cancer–related death factor is Ovarian Cancer (OC) which is regarded as the fifth leading cause of cancer mortality in women [1, 2]. OC is diagnosed in approximately 62% of cases in advanced stages (III-IV). The most lethal and common type of OC is high-grade serous ovarian cancer (HGSOC). Debulking surgery and systemic therapy are the standard treatments for OC. The best survival rates are obtained when tumor implants are completely resected; postsurgical residual tumor is a significant adverse prognostic factor. Thus, it is essential to detect the tumor spread sites. Almost 70–80% of OC patients show recurrence of the disease despite proper first-line chemotherapy and optimal surgery. Recurrence occurs in 23% of cases within or during 6 months following the ending of primary chemotherapy, and it is 60% after 6 months [3–5]. Therefore, it is required to apply an accurate noninvasive follow-up method to follow-up patients with OC. One of the useful approaches in OC post-treatment surveillance is Fluorine-18 fluorodeoxyglucose positron emission tomography / computed tomography (18F-FDG PET/CT) in comparison to conventional modalities, such as magnetic resonance imaging (MRI) or computed tomography (CT) [6–9]. Distant metastases, recurrent disease, and staging are detected, and the response to therapy is monitored using FDG PET/CT in OC [7–12].

Currently, it is required to interpret FDG PET/CT images in clinical practice by nuclear medicine physicians. With the increasing burden of image interpretation, the oversight risk or misdiagnosis increases. Hence, an automated system is demanded to avoid such problems. One of the artificial intelligence (AI) methods in the medical field is image analysis which uses convolutional neural networks (CNNs) and is regarded as a machine learning (ML) approach. CNN, as a subdivision of deep neural network (deep learning-DL) methods, is a beneficial instrument for diagnostic imaging [13]. Researchers have developed various CNNs for diagnostic imaging with outstanding performance in image processing activities, such as classification, detection, segmentation, and registration using radiography, ultrasound (US), MRI, and CT [14–16]. Besides, CNN methods have shown successful outcomes with PET/CT [17, 18]. Wang et al. used CT images by a CNN to predict OC recurrence [19]. Moreover, Bogani et al. used the data set of patients with recurrent OC to develop a DL model for the analysis of the factors that predict complete cytoreduction during secondary cytoreductive surgery [20]. These studies have used 2-dimensional (2D) convolutional neural network (CNN)-based DL models to investigate OC recurrence prediction. Although there are a large number of studies on DL-based medical image analysis usage, it should be mentioned that there are a few studies that directly exploited the 3D tomographic medical imaging data [21, 22].

Additionally, 3-dimensional (3D) convolutional neural network (CNN)-based DL models offered considerable benefits in voxel-wise analysis, combining information from the 2D plane domain and expanding the voxel spatial information perception of the model [23]. Several studies have revealed that outstanding diagnostic results have been achieved by 3D CNN in different imaging modalities and cancer types [22–25]. To the best of our knowledge, there are few studies investigating the recurrence/post-therapy surveillance for OC patients by 18F-FDG PET/CT, and 3D CNN approaches. It was assumed that human misdiagnosis/oversight would be prevented by presenting an automated system for following up on recurrence/post-therapy. Furthermore, such as system is beneficial for radiologists in selecting patients that need urgent interpretation. Notably, this system is helpful for physicians and physicists who are inexperienced in nuclear medicine.

In this study, we aimed to develop the 3D CNN-based diagnosis system for recurrence/post-therapy surveillance with the capability of classifying whole-body FDG PET images into two classes: 1) cancerous and 2) non-cancerous. We also assessed the diagnostic effectiveness of the proposed system for the whole-body tumor staging of OC patients. Using this system, physicians can perform radiology-based diagnoses to double-check their opinions.

Patients

The research ethics committees of Shiraz University approved this study (approval ID: IR.US.REC.1401.001). The following were the inclusion criteria: (a) Patients with histopathology diagnosis of HGSOC; (b) 18F-FDG PET/CT scan acquired as per the protocol followed at Kowsar hospital between April 2019 and May 2022; (c) Stage III or IV with or without known residual disease after initial treatment; (d) Unexpected rise of Cancer Antigen 125 (CA125) with negative or suspicious conventional imaging. The following were the exclusion criteria: (a) Uncontrolled diabetic patients; (b) Recent history (less than five years) of other malignancy. We anonymized all data sets before the analysis. Table 1 presents the demographic information of patients. Figure 1 indicates a flowchart of the process of patient selection.

Table 1

Patient characteristics
Characteristics		n (%)
No. of patients		37
Age (y)	Mean ± SD	56.3 ± 10.3
Age (y)	Range	36–83
Diagnostic classification	Non-Cancerous	9 (24)
Diagnostic classification	Cancerous	28 (76)
Stage	I	0
	II	0
	III	9 (32)
	IV	19 (68)

For recurrence/post-therapy surveillance, diagnostic classification was determined by PET/CT images and pathological findings. Also, the classification system of the International Federation of Gynecology and Obstetrics (FIGO) 2014 was used for staging patients [26] by the use of pathological and surgical diagnosis and PET/CT images. Besides the pathological findings, a nuclear medicine physician with 26 years of PET/CT image interpretation experience and an experienced nuclear medicine physicist assessed the image sets independently.

PET/CT Protocol

After treatment, all patients underwent 18F-FDG PET/CT imaging. Following intravenous administration of 190–410 MBq (5.1–11 mCi) 18F-FDG, patients rested in a quiet room, and then whole-body 18F-FDG PET/CT images were acquired. All 18F-FDG PET/CT examinations were performed after 60 minutes of rest by the use of the Philips Ingenuity TF 64 PET/CT (Philips Medical Systems). Table 2 shows the clinical parameters.

Table 2

Clinical parameters of the Philips Ingenuity TF 64 PET/CT
CT scanning
Tube voltage	120 kV
Effective tube	100 mA (variable by Dose Right)
Matrix	512 × 512×256–340
Detector configuration	64×0.625 mm
Slice thickness, mm	2
Transverse FOV, mm	600
PET scanning
18F-FDG injection dose, MBq/kg	3.7
Scan time for each bed, mm	90
TOF	Yes
PET reconstruction
Reconstruction	3D-OSEM
Iterations	3
Subsets	33
Smoothing	N/A
FWHM of filter, mm	4.8
Matrix	144×144×192–255
Pixel size, mm	4×4×4

Data Set

Two categories were considered for detecting the most appropriate 3D CNN for recurrence/post-therapy surveillance: diagnostic classification and staging. The classification of the PET/CT images were done as non-cancerous and cancerous for creating a dataset. This classification was done based on the pathological findings as well as the consensus of a nuclear medicine physician and a nuclear medicine physicist. A total of 1224 image sets were allocated randomly to the training and test groups. 1054 image sets (augmented images) were used in the training phase from 37 patients (9 patients in the non-cancerous group and 28 in the cancerous group). 170 image sets (augmented images) were used in the test phase from 10 patients (5 in the non-cancerous group and 5 in the cancerous group).

Two stages (III-IV) as the most prevalent ones in patients with OC were considered for evaluating the diagnostic utility of the 3D CNN for the characterization of whole-body tumor staging. Pathological and surgical diagnosis, as well as the consensus of a nuclear medicine physician and a nuclear medicine physicist, were used for determining the patients’ FIGO staging. A total of 1224 image sets were allocated randomly to the training and test groups. 1054 image sets (augmented images) were utilized from 28 patients (9 ones in the stage III group and 19 ones in the stage IV group) in the training phase. 170 image sets (augmented images) were utilized from 10 patients (5 ones with stage III and 5 ones with stage IV) in the test phase. The section assigned to image preprocessing explains the augmentation method in detail.

Image Preprocessing

In this phase, the codes written in the programming language Python 3.9.12 were used to process and augment the images. The whole-body scans were in Digital Imaging and Communications in Medicine format (DICOM). DICOM image files and the related data are composed of 2D layers. Thus, the DICOM images were converted into Neuroimaging Informatics Technology Initiative (NIfTI) images that stored images and other data in a 3D format. The margins, which contained the patient information, were resized and cropped to 128 × 128 × 128 pixels for all image sets. Images were processed for the training image sets by cutting out the bottom and top of patient images, nearly corresponding to the legs and brain. The data augmentation was done in two steps: first, the images were flipped and rotated at limited angles in image preprocessing (offline augmentation). Then, rotation was done with more angles during the 3D CNN model training (on-the-fly augmentation). The bottom and top of the images (nearly corresponding legs and brain) were cut out for the validation and testing image sets, similar to the training phase set. Rotation and flipping were performed to augment the images in image preprocessing. Images were normalized for preparing the data sets to enter the 3D CNN models. Given that PET is a relatively noisy process in comparison with other imaging modalities [27], a 10% Gaussian denoising filter is applied to enhanced diagnostic performance. The schematic of preprocessing steps is shown in Fig. 2. In this study, CT images were used merely for attenuation correction, not for staging and classification. From each of the augmented images, output with the NIfTI format was obtained. Subsequently, all of these images underwent thorough review and validation by a nuclear medicine physicist.

The OCDAc-Net and OCDAs-Net architectures

The OCDA-Net in this study is a 3D CNN-based on ResNet-50 architecture. ResNet-50 has 50 layers, and the model contains above 23 million trainable parameters, implying a deep architecture more suitable for image classification. The small word after OCDA indicates that the network is used for diagnostic classification (added c) or staging (added s). The OCDA-Net has over 46 million trainable parameters, and we used this to classify whole-body FDG PET images into 1) cancerous and 2) non-cancerous classes and tumor staging of patients into 1) stage III and 2) stage IV. The details on 3D CNN architectures and the methods applied in the present study are explained here. The 3D CNN architecture included eight Residual connections. This network uses two special Residual blocks, one of which contains an additional layer on its skip connection and is called nResidual. The structure of "Residual" is as follows. “1. Convolutional layer 1, 2. Batch normalization 1, 3. Activation layer 1 (ReLU), 4. Convolutional layer 2, 5. Batch normalization 2, 6. Activation layer 2 (ReLU), 7. Convolutional layer 3, 8. Batch normalization 3, 9. Merge layer (Add), 10. Activation layer 3 (ReLU)”. The structure of "nResidual" is as follows. “1. Convolutional layer 1, 2. Batch normalization 1, 3. Activation layer 1 (ReLU), 4. Convolutional layer 2, 5. Batch normalization 2, 6. Activation layer 2 (ReLU), 7. Convolutional layer 3, 8. Batch normalization 3, 9. Convolutional layer 4, 10. Batch normalization 4, 11. Merge layer (Add), 12. Activation layer 3 (ReLU)”. Table 3 and Fig. 3 show this architecture in detail. Feature maps are created by 3D Convolution layers, which extract image features. 3D Pooling layers can reduce the data amount and enhance the robustness against misregistration by reducing of size (downsampling) in the resulting feature map. “Residual” is a block, as a ResNet feature, which makes the combination of the layers and can solve the problem of conventional gradient disappearance. The neurons in a layer are connected to the related neurons in the previous layer. Also, this network applies batch normalization, sigmoid layers, and a rectified linear unit function (ReLU).

Table 3

Details of the OCDA-Net architecture
Layer	Output Size	Repeat Count
Input Layer	(128, 128, 128, 1)
Zero Padding 3D	(134, 134, 134, 1)	1
Convolutional 3D	(64, 64, 64, 64)	1
Batch Normalization	(64, 64, 64, 64)	1
Activation	(64, 64, 64, 64)	1
Max Pooling 3D	(31, 31, 31, 64)	1
nResidual 1	(31, 31, 31, 256)	1
Residual 2	(31, 31, 31, 256)	2
nResidual 3	(16, 16, 16, 512)	1
Residual 4	(16, 16, 16, 512)	3
nResidual 5	(8, 8, 8, 1024)	1
Residual 6	(8, 8, 8, 1024)	5
nResidual 7	(4, 4, 4, 2048)	1
Residual 8	(4, 4, 4, 2048)	2
Average Pooling 3D	(2, 2, 2, 2048)	1
Fully Connected	(1)
Total Parameters: 46,147,993 Trainable Parameters: 46,146,445 Non-trainable Parameters: 1,548

The image sets were split into three subsets: training (75% of all patients), validation (10%), and testing (15%). On-the-fly augmentation was employed as random rotations for training sets. Using the validation set, all model hyperparameters (for example, network architecture, decay steps, and learning rate) were selected, and the best-trained 3D CNN was determined. Using the test subset, the final performance of the 3D CNN was assessed. The networks were trained for 100 epochs with a batch size of 2 using an Adam optimizer with an initial learning rate of 0.00001 in the OCDAc-Net and 0.000001 in the OCDAs-Net with exponential decay of initial learning rate and decay steps of 100000, decay rate of 0.99 and binary cross-entropy loss, by which point the validation accuracy had stopped increasing.

In order to build DL models, we used the Python programming language, TensorFlow 2.9.1, and Keras 2.9.0 framework for neural networks. A computer with a Core i7-11700K 3.60-GHz central-processing unit, 64 GB of random-access memory, and a GeForce RTX 3090 graphics-processing unit was used to perform DL.

Statistical Analysis

For evaluating the 3D CNN-based diagnosis system performance, the test data set is utilized for calculating the overall accuracy, sensitivity (recall), precision, confusion matrix, and F-measure (F1-Score) for classification and staging. In our 3D CNN, the output of the classification into non-cancerous and cancerous groups was a continuous number between 0 to 1: 0.50–1.00 was regarded as cancerous, and 0–0.49 was considered non-cancerous same as staging for stage IV and stage III, respectively. Receiver operating characteristic (ROC) analysis was conducted to obtain the area under the ROC (AUC) to evaluate the diagnostic performance of the 3D CNN models. A bootstrap of 1000 random resamples was used to report the 95% confidence interval (CI). Python 3.9.12 was used in this study for statistical analysis.

A total of 37 women (age range, 36–83 years; mean age, 56.3 years) were assessed in the data sets. There were 9 patients classified as non-cancerous and 28 patients classified as cancerous, considered as diagnostic classification; out of 28 cancerous cases, 19 patients were classified as stage IV, and 9 patients were classified as stage III according to the pathological findings and opinion of physicians considered as staging. 1224 image sets were obtained after augmentation for diagnostic classification and staging separately. 612 (training, 527; testing, 85) image sets in the cancerous group and 612 (training, 527; testing, 85) image sets in the non-cancerous group were utilized for the diagnostic classification in the network. 612 (training, 527; testing, 85) image sets in stage III and 612 (training, 527; testing, 85) image sets in stage IV were utilized for the staging phase in the network. Figure 4 shows the performance of the proposed OCDAc-Net and OCDAs-Net as a confusion matrix. Overall accuracy, recall, precision, and F-measure are calculated with the following formulas (1–4).

$$\text{A}\text{c}\text{c}\text{u}\text{r}\text{a}\text{c}\text{y}=\frac{\text{N}\text{o}. \text{o}\text{f} \text{i}\text{m}\text{a}\text{g}\text{e}\text{s} \text{c}\text{o}\text{r}\text{r}\text{e}\text{c}\text{t}\text{l}\text{y} \text{c}\text{l}\text{a}\text{s}\text{s}\text{i}\text{f}\text{i}\text{e}\text{d}}{\text{T}\text{o}\text{t}\text{a}\text{l} \text{n}\text{o}. \text{o}\text{f} \text{i}\text{m}\text{a}\text{g}\text{e}\text{s}} \left(1\right)$$

$$\text{P}\text{r}\text{e}\text{c}\text{i}\text{s}\text{i}\text{o}\text{n}=\frac{\text{S}\text{u}\text{m} \text{o}\text{f} \text{a}\text{l}\text{l} \text{T}\text{r}\text{u}\text{e} \text{P}\text{o}\text{s}\text{i}\text{t}\text{i}\text{v}\text{e}\text{s} \left(\text{T}\text{P}\right)}{\text{S}\text{u}\text{m} \text{o}\text{f} \text{a}\text{l}\text{l} \text{T}\text{r}\text{u}\text{e} \text{P}\text{o}\text{s}\text{i}\text{t}\text{i}\text{v}\text{e}\text{s} \left(\text{T}\text{P}\right)+\text{A}\text{l}\text{l} \text{F}\text{a}\text{l}\text{s}\text{e} \text{P}\text{o}\text{s}\text{i}\text{t}\text{i}\text{v}\text{e}\text{s} \left(\text{F}\text{P}\right)} \left(2\right)$$

$\text{R}\text{e}\text{c}\text{a}\text{l}\text{l}=\frac{\text{S}\text{u}\text{m} \text{o}\text{f} \text{a}\text{l}\text{l} \text{T}\text{r}\text{u}\text{e} \text{P}\text{o}\text{s}\text{i}\text{t}\text{i}\text{v}\text{e}\text{s} \left(\text{T}\text{P}\right)}{\text{S}\text{u}\text{m} \text{o}\text{f} \text{a}\text{l}\text{l} \text{T}\text{r}\text{u}\text{e} \text{P}\text{o}\text{s}\text{i}\text{t}\text{i}\text{v}\text{e}\text{s} \left(\text{T}\text{P}\right)+\text{A}\text{l}\text{l} \text{F}\text{a}\text{l}\text{s}\text{e} \text{N}\text{e}\text{g}\text{a}\text{t}\text{i}\text{v}\text{e}\text{s} \left(\text{F}\text{N}\right)} \left(3\right)$

$\text{F}-\text{m}\text{e}\text{a}\text{s}\text{u}\text{r}\text{e}=\frac{2\text{}\text{P}\text{r}\text{e}\text{c}\text{i}\text{s}\text{i}\text{o}\text{n}\text{}\text{R}\text{e}\text{c}\text{a}\text{l}\text{l}}{\text{P}\text{r}\text{e}\text{c}\text{i}\text{s}\text{i}\text{o}\text{n}+\text{R}\text{e}\text{c}\text{a}\text{l}\text{l}} \left(4\right)$

Table 4 shows the overall accuracy and AUC of the OCDAc-Net and the OCDAs-Net for diagnostic classification and staging. These performance metrics, as the top metrics, are employed for measuring the classification algorithms’ performance [28]. With the fully optimized 3D CNN, the OCDAc-Net reached an overall accuracy of 92%, and the recall, precision, and F-measure were 99, 88 and 93% for the non-cancerous class, respectively. The cancerous class showed a relatively lower performance than the other class except for the precision which contributed to the lower accuracy for diagnostic classification.

Table 4

Performance of the OCDAc-Net and the OCDAs-Net for diagnostic classification and staging
Class	Overall Accuracy	AUC	95% CI
OCDAc-Net	0.92	0.990	0.983–0.997
OCDAs-Net	0.94	0.995	0.990–0.999

After giving images of non-cancerous and cancerous patients to the learned model, 98.8 and 85.8% accuracy were obtained, respectively. The fully optimized CNN, the OCDAs-Net achieved an overall accuracy of 94%, with the recall, precision, and F measure as 95, 92 and 94% for stage III, respectively. Similar to the diagnostic classification, the precision is lower than stage IV for stage III. When the learned model was applied, 95.3% of patients with stage III and 91.7% with stage IV were correctly predicted. Table 5 presents the networks’ class-wise performance. The 3D CNNs showed the capability to predict diagnostic classification and staging by the experts into PET images, which were cancerous versus non-cancerous and stage III versus stage IV with an AUC of 0.990 (95% CI: 0.983, 0.997) and 0.995 (95% CI: 0.990, 0.999), respectively. The ROC curves are shown in Fig. 5.

Table 5

The class-wise performance of the OCDAc-Net and the OCDAs-Net
Category	Class	Precision	Recall	F-measure
Diagnostic classification	Cancerous	0.99	0.86	0.92
Diagnostic classification	Non-cancerous	0.88	0.99	0.93
Staging	Stage III	0.92	0.95	0.94
Staging	Stage IV	0.95	0.92	0.93

In the present study, we developed 3D CNN-based models named the OCDAc-Net and the OCDAs-Net using DL to predict automatic diagnostic classification and staging of OC patients with routine clinical PET/CT data. A review of clinical symptoms was recommended by the Society of Gynecologic Oncology (SGO), along with examining patients physically, which needs to be performed after the initial treatment for OC with radiographic imaging (CT, MRI, or PET) and an optional CA125 test every three months in patients with suspected recurrence. These patients commonly have an incremented level of CA125 and results on clinical examination or suspicious symptoms [29]. More accurate information can be provided by 18F-FDG PET/CT examination on surveillance and staging to detect recurrent high-grade OC. Though, surgical staging is still conducted owing to its higher false negative rate for micro lesions or cystic lesions [9].

DL has been used in a few studies on gynecological imaging. A systematic review was performed by Akazawa and Hashimoto for elucidating the state of AI studies on gynecologic cancers. They concluded that cervical cancer had been studied highly compared to endometrial and OC. Moreover, prognoses were used mostly in the investigation of cervical cancer, while diagnoses were used primarily to study OC [17]. Recently, an automated framework was constructed by Schwartz et al. to identify OC in transgenic mice based on optical coherence tomography (OCT) recordings [24]. A two-stage supervised 3D CNN was developed by Potočnik and Šavc based on the U-Net established to automate detecting ovarian follicles in ultrasound images [25].

Several other AI algorithms exist for PET image detection and classification. A DL model was constructed by Shen et al. to forecast distant metastasis and local relapse after concurrent chemoradiotherapy (CCRT). Distant and local failures were experienced by 26 and 21 patients between 142 patients with cervical cancer in the median 40 months of follow-up, respectively. The distant metastasis and local relapse were predicted by their model from the PET/CT images with accuracies of 0.87 and 0.89, respectively [30]. A system based on CNN was developed by Kawauchi et al. to categorize whole-body FDG PET as 1) benign, 2) equivocal, or 3) malignant. They classified 1280 (37%), 755 (22%), and 1450 (42%) patients as benign, equivocal, and malignant, respectively. In the analysis results, benign, equivocal, and malignant images were predicted by CNN with accuracies of 0.99, 0.87, and 0.99, respectively [31].

To our knowledge, this is the first staging and image classification study for recurrence/post-therapy surveillance of OC utilizing PET images via DL. It should be noted that the predictions were made for each patient using 3D images. We reported the final performance of the models. In diagnostic classification, the cancerous and non-cancerous classes were correctly predicted by the neural network with an overall accuracy of 92%. However, the accuracy was 85.8% for cancerous patients. Thus, according to the average probability of all metrics, 3D CNN can effectively predict 2-class classification from PET images. Moreover, it was classified correctly with an accuracy of 95.3% for stage III, and 91.7%, for stage IV for predicting the staging with an overall accuracy of 94%. It was suggested that the system might potentially assist radiologists in preventing misdiagnosis and oversight. The automatic 3D CNN-based classification and staging has sufficient accuracy. Hence, valuable assistance can be provided by our approach for PET/CT readers in the diagnosis and staging of patients with OC. Although the process is automated, the physician expert should still supervise the final report. Thus, the possibility of assisted reading is offered by our approach, which may help PET/CT readers in training. As a double-checking system, the current system could help physicians. The 3D CNN would also present experience from experts performing 18F-FDG PET/CT regularly for gynecological diseases. Inter-observer agreements could be increased between institutions (such as clinical trials).

It should be mentioned that when CNN directly learns with 3D images, the computational complexity becomes enormous [31]. The lack of data is typically the key factor causing 3D CNN-based detection approaches to perform less efficiently [25]. However, we proved that an effective 3D model can be developed through supervised learning appropriately, in spite of the relatively small training set. The accuracy could be further improved by some approaches. In the present study, a network based on ResNet-50 was used to decrease the learning cost [32], a type of the “ResNet” family. Indeed, ResNet systems can be technically built with deeper layers. Recently, different networks have been developed based on ResNet with higher performance [33, 34]. Considering the big-data science, the number of images must increase for further enhancement in diagnostic accuracy [35]. Using data augmentation methods, we increased the data sets in this study.

There were several limitations to this study. First, a black box was essentially the working mechanism of the DL algorithm. The feature extraction and selection processes are opaque and automated. This is a general problem for physicians to understand the model results intuitively and clearly. Therefore, it is challenging to determine the reasoning behind this phenomenon and the negative or positive effects of certain characteristics [36]. Second, our prediction model purely focused on PET images and could not benefit potentially from a larger cohort of data sets. Third, in the present study, we utilized only the data sets of patients with stages III and IV according to the review of recurrence/post-therapy surveillance and available data. In the future, it is essential to perform a large-scale, prospective study utilizing various 18F-FDG-PET/CT data.

To conclude, the 3D CNN-based models were proposed in this study to predict diagnostic classification and staging among patients with OC. It aimed to provide a potential noninvasive tool to guide clinical decisions. It was indicated that DL can provide new models related to the recurrence/post-therapy surveillance of OC by clinical PET/CT imaging, which demonstrated promising prognostic value versus clinical characteristics. An acceptable operating point will be identified in future work toward clinical implementation while expanding and validating in larger independent data sets.

Acknowledgements

This work was supported by the International Atomic Energy Agency (IAEA) under the coordinated research project (CRP) E13050 - the International Multicenter Trial on FDG PET/CT in Ovarian Cancer (POCA). The authors would like to acknowledge the instrumental and technical support of the Nuclear Medicine Department, Kowsar Hospital.

No potential conflicts of interest were disclosed.

Funding:

This work was supported by Shiraz University (Grant numbers 1GBC1M148325). Also, the work was funded by the International Atomic Energy Agency (IAEA) under the coordinated research project (CRP) E13050 - the International Multicenter Trial on FDG PET/CT in Ovarian Cancer (POCA).

Dondi F, Albano D, Bertagna F, Giubbini R. [18F] FDG PET/CT and CA-125 in the evaluation of ovarian cancer relapse or persistence: is there any correlation? Nuclear Med Rev. 2022;25(2):78–84.
Ovarian Cancer. — Cancer Stat Facts [Internet]. [cited 2023 Jan 2]. Available from: https://seer.cancer.gov/statfacts/html/ovary.html.
Perrone AM, Dondi G, Lima GM, Castellucci P, Tesei M, Coluccelli S, et al. Potential prognostic role of 18F-FDG PET/CT in invasive epithelial ovarian cancer relapse. A preliminary study. Cancers. 2019;11(5):713.
Wang J, Liu L, Pang H, Liu L, Jing X, Li Y. Preoperative PET/CT score can predict incomplete resection after debulking surgery for advanced serous ovarian cancer better than CT score, MTV, tumor markers and hematological markers. Acta Obstet Gynecol Scand. 2022;101(11):1315–27.
Salvi A, Hardy LR, Heath KN, Watry S, Pergande MR, Cologna SM et al. PAX8 modulates the tumor microenvironment of high grade serous ovarian cancer through changes in the secretome. Neoplasia [Internet]. 2023;36:100866. Available from: https://www.sciencedirect.com/science/article/pii/S1476558622000914.
Rusu G, Achimaș-Cadariu P, Piciu A, Căinap SS, Căinap C, Piciu D. In: Healthcare, editor. A Comparative Study between 18F-FDG PET/CT and Conventional Imaging in the Evaluation of Progressive Disease and Recurrence in Ovarian Carcinoma. Multidisciplinary Digital Publishing Institute; 2021. p. 666.
Virarkar M, Ganeshan D, Gulati AT, Palmquist S, Iyer R, Bhosale P. Diagnostic performance of PET/CT and PET/MR in the management of ovarian carcinoma—a literature review. Abdom Radiol. 2021;46(6):2323–49.
Mallum A, Mkhize T, Akudugu JM, Ngwa W, Vorster M. The Role of Positron Emission Tomography and Computed Tomographic (PET/CT) Imaging for Radiation Therapy Planning: A Literature Review. Diagnostics. 2022;13(1):53.
Pillay V, Hadebe B, Vorster M. Molecular Imaging and Theranostics in Ovarian Cancer: The Role of Nuclear Medicine. Exon Publications. 2022;69–85.
Tsuyoshi H, Yoshida Y. Diagnostic imaging using positron emission tomography for gynecological malignancy. J Obstet Gynecol Res. 2017;43(11):1687–99.
Kim SJ, Lee SW. Diagnostic accuracy of 18F-FDG PET/CT for detection of peritoneal carcinomatosis; a systematic review and meta-analysis. Br J Radiol. 2018;91(1081):20170519.
Vahidfar N, Farzanefar S, Ahmadzadehfar H, Molloy EN, Eppard E. A Review of Nuclear Medicine Approaches in the Diagnosis and the Treatment of Gynecological Malignancies. Cancers. 2022;14(7):1779.
Hirata K, Sugimori H, Fujima N, Toyonaga T, Kudo K. Artificial intelligence for nuclear medicine in oncology. Annals of Nuclear Medicine [Internet]. 2022;36(2):123–32. Available from: https://doi.org/10.1007/s12149-021-01693-6.
Saba L, Biswas M, Kuppili V, Godia EC, Suri HS, Edla DR, et al. The present and future of deep learning in radiology. Eur J Radiol. 2019;114:14–24.
Satoh Y, Imokawa T, Fujioka T, Mori M, Yamaga E, Takahashi K et al. Deep learning for image classification in dedicated breast positron emission tomography (dbPET). Annals of Nuclear Medicine [Internet]. 2022;36(4):401–10. Available from: https://doi.org/10.1007/s12149-022-01719-7.
Sharma G, Prabha C. A systematic review for detecting cancer using machine learning techniques. In: AIP Conference Proceedings. AIP Publishing LLC; 2022. p. 40007.
Akazawa M, Hashimoto K. Artificial intelligence in gynecologic cancers: Current status and future challenges–A systematic review. Artif Intell Med. 2021;120:102164.
Saba T. Recent advancement in cancer detection using machine learning: Systematic survey of decades, comparisons and challenges. J Infect Public Health. 2020;13(9):1274–89.
Wang S, Liu Z, Rong Y, Zhou B, Bai Y, Wei W, et al. Deep learning provides a new computed tomography-based prognostic biomarker for recurrence prediction in high-grade serous ovarian cancer. Radiother Oncol. 2019;132:171–7.
Bogani G, Rossetti D, Ditto A, Martinelli F, Chiappa V, Mosca L et al. Artificial intelligence weights the importance of factors predicting complete cytoreduction at secondary cytoreductive surgery for recurrent ovarian cancer. J Gynecologic Oncol. 2018;29(5).
Avanzo M, Porzio M, Lorenzon L, Milan L, Sghedoni R, Russo G, et al. Artificial intelligence applications in medical imaging: A review of the medical physics research in Italy. Physica Med. 2021;83:221–41.
Trivizakis E, Manikis GC, Nikiforaki K, Drevelegas K, Constantinides M, Drevelegas A, et al. Extending 2-D convolutional neural networks to 3-D for advancing deep learning cancer classification with application to MRI liver tumor differentiation. IEEE J biomedical health Inf. 2018;23(3):923–30.
Li X, Gao H, Zhu J, Huang Y, Zhu Y, Huang W, et al. 3D Deep Learning Model for the Pretreatment Evaluation of Treatment Response in Esophageal Carcinoma: A Prospective Study (ChiCTR2000039279). Int J Radiation Oncology* Biology* Phys. 2021;111(4):926–35.
Schwartz D, Sawyer TW, Thurston N, Barton J, Ditzler G. Ovarian cancer detection using optical coherence tomography and convolutional neural networks. Neural Comput Appl. 2022;34(11):8977–87.
Potočnik B, Šavc M. Deeply-Supervised 3D Convolutional Neural Networks for Automated Ovary and Follicle Detection from Ultrasound Volumes. Appl Sci. 2022;12(3):1246.
Mutch DG, Prat J. 2014 FIGO staging for ovarian, fallopian tube and peritoneal cancer. Gynecol Oncol. 2014;133(3):401–4.
Schaefferkoetter J, Yan J, Ortega C, Sertic A, Lechtman E, Eshet Y, et al. Convolutional neural networks for improving image quality with noisy PET data. EJNMMI Res. 2020;10(1):1–11.
Khan AI, Shah JL, Bhat MM, Coronet. A deep neural network for detection and diagnosis of COVID-19 from chest x-ray images. Comput Methods Programs Biomed. 2020;105581.
Matulonis UA, Sood AK, Fallowfield L, Howitt BE, Sehouli J, Karlan BY. Ovarian cancer. Nat reviews Disease primers. 2016;2(1):1–22.
Shen WC, Chen SW, Wu KC, Hsieh TC, Liang JA, Hung YC, et al. Prediction of local relapse and distant metastasis in patients with definitive chemoradiotherapy-treated cervical cancer by deep learning from [18 F]-fluorodeoxyglucose positron emission tomography/computed tomography. Eur Radiol. 2019;29(12):6741–9.
Kawauchi K, Furuya S, Hirata K, Katoh C, Manabe O, Kobayashi K, et al. A convolutional neural network-based system to classify patients using FDG PET/CT examinations. BMC Cancer. 2020;20(1):1–10.
He K, Zhang X, Ren S, Sun J. Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. 2016. p. 770–8.
Iandola FN, Han S, Moskewicz MW, Ashraf K, Dally WJ, Keutzer K. SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and < 0.5 MB model size. arXiv preprint arXiv:160207360. 2016.
Zagoruyko S, Komodakis N. Wide residual networks. arXiv preprint arXiv:160507146. 2016.
Matsubara K, Ibaraki M, Nemoto M, Watabe H, Kimura Y. A review on AI in PET imaging. Annals of Nuclear Medicine [Internet]. 2022;36(2):133–43. Available from: https://doi.org/10.1007/s12149-021-01710-8.
Xu Y, Hosny A, Zeleznik R, Parmar C, Coroller T, Franco I, et al. Deep learning predicts lung cancer treatment response from serial medical imaginglongitudinal deep learning to track treatment response. Clin Cancer Res. 2019;25(11):3266–75.

Download PDF

Journal Publication

published 28 Sep, 2023

Read the published version in Annals of Nuclear Medicine →

Reviewers agreed at journal
27 Jun, 2023
Reviewers invited by journal
27 Jun, 2023
Editor assigned by journal
26 Jun, 2023
First submitted to journal
23 Jun, 2023

You are reading this latest preprint version

The OCDA-Net: a 3D convolutional neural network-based system for classification and staging of ovarian cancer patients using FDG PET/CT examinations

Status:

Journal Publication

Version 1

Abstract

Objective

Methods

Results

Conclusions

Figures

Introduction

Materials and Methods

Patients

PET/CT Protocol

Data Set

Image Preprocessing

The OCDAc-Net and OCDAs-Net architectures

Statistical Analysis

Results

Discussion

Conclusion

Declarations

References

Status:

Journal Publication

Version 1