Ovarian Cancer (OC) is the most lethal gynaecologic malignancy worldwide, ranking as the fifth deadliest cancer among women and accounting for approximately 13000 deaths in 2023 in the United States [1]. According to guidelines, suspected OC patients firstly undergo pelvic ultrasound, Computed Tomography of the Thorax, Abdomen and Pelvis (CT TAP) and CA125 measurement for staging purposes. Depending on the CT TAP results and clinical assessment, clinicians evaluate the tumor resectability. Patients likely to achieve complete tumor resection undergo primary debulking surgery followed by adjuvant chemotherapy. Otherwise, they receive neoadjuvant chemotherapy, followed by interval debulking surgery and adjuvant chemotherapy.
Although most patients initially respond positively to this standard of care, it is estimated that 70% of patients will experience a relapse. Surgical intervention aims at achieving complete tumor resection; however, it often results to be either aggressive, leading to severe postoperative complications, or ineffective, resulting in incomplete tumor removal and an associated twofold increase in the risk of death, with the latter scenario occurring in approximately 40% of cases [2]. The challenge in clinical practice is accurately predicting the success of cytoreductive surgery, critical due to the severe consequences of misjudgment, such as unnecessary invasive procedures causing significant perioperative complications and emotional distress. The complexity of predicting surgical outcomes is heightened by the varied and distinct presentations of OC -four clinical cases are shown in Figure 1 -making it difficult to assess tumor resectability from diagnostic imaging. Advancements in this area are crucial to minimize unnecessary surgeries and tailor treatments to patient-specific needs. Nowadays radiomics, a computational tool for extracting high-dimensional features from medical images, becomes part of personalized oncology treatments, driven by advancements in Machine Learning (ML) [3]. However, ML requires appropriate selection among the numerous radiomic features extracted from images [4] [5]. Deep Learning (DL) has shown promising results in automatically and directly detecting valuable features from medical imaging [6], boosting the progress of computer vision in the medical field [7] [8], and demonstrating superior performance in comparison to hand-crafted image features [9]. In this paper, a 3D CNN was designed to perform binary TR classification of patients with OC. Specifically, what emerged from literature research was the absence of robust radiological indexes to select patients for total surgical resection. Hence, our primary contribution was the implementation of a non-invasive and preoperative DL model to assess whether an upfront patient could be a suitable candidate for debulking surgery, when achieving a total resection appears feasible, or the patient might be recommended to undergo neoadjuvant chemotherapy before proceeding to interval surgery, when a complete resection seems unlikely. Therefore, the proposed model might potentially assist radiologists and gynecologists to assess TR and guide therapeutic strategies for patients with OC.
Artificial Intelligence (AI) has been demonstrated to enhance the effectiveness of tumor detection, classification, and treatment monitoring in cancer imaging.The integration of radiomics and DL enabled the extraction of image features and information, which might be imperceptible to the human subjective evaluations, yielding to promising medical applications [10]. In the context of OC, several noteworthy studies have been conducted. In the domain of radiomics and ML domain, Lu et al. [11] proposed an approach to predict two-year overall survival in 364 epithelial OC patients. In this study, 657 quantitative descriptors were extracted from preoperative CT images, upon which the ML algorithm Radiomic Prognostic Vector was developed. The latter accurately identified the 5% of patients with a median overall survival of less than 2 years, demonstrating significant improvement over established prognostic methods. Crispin-Ortuzar et al. [12] addressed the challenge of predicting neoadjuvant chemotherapy (NACT) response of 72 highgrade serous OC (HGSOC) affected patients, presenting an ensemble ML model that, integrating baseline clinical, blood-based, and radiomic biomarkers from primary and metastatic lesions, predicted changes in total disease volume. Validation on internal and external cohorts showed that the model significantly improved prediction accuracy compared to the clinical model, highlighting the potential of radiomics in enhancing treatment response predictions. On the front of DL, Jan et al. [10] developed an AI ensemble model combing radiomics, DL and clinical features from CT images to distinguish between benign and malignant OC. With 149 patients and 185 tumors, the model achieved 82% accuracy, 89% specificity, and 68% sensitivity. Compared to junior radiologists, the model exhibited higher accuracy and specificity while maintaining comparable sensitivity. Wang et al. [7] proposed a DL method to predict 3-year recurrence in 245 high-grade serous OC patients from preoperative CT images. The DL network, trained on 8917 CT images, extracts a 16-dimensional DL feature used to predict the outcome probability. The model achieved AUC values of 0.772 and 0.825 for high and low recurrence risk, exhibiting stronger prognostic value compared to clinical characteristics. Zheng et al. [13] proposed a Vit-based DL model for predicting overall survival in 734 high-grade serous OC patients using preoperative CT images. Analyzing 734 patients, the dataset was split into training (n = 550) and validation (n = 184) cohorts. The model demonstrated robust performance with AUC = 0.822 in the training cohort of 550 patients and AUC = 0.823 in the validation cohort of 184 women. Lei et al. [14] developed a DL model for predicting platinum sensitivity in 93 patients with epithelial OC using contrast-enhanced magnetic resonance imaging (MRI). A pre-trained CNN were used and 1,024 features were automatically extracted from MRI sequences to predict platinum sensitivity.The model performed Area Under the Curve (AUC) of 0.97 and 0.98 in training and validation cohorts. Among the 20 research papers examining OC in [15], 11 primarily aimed at classifying between benign, malignant, and/or borderline tumors. Two of these studies focused on resistance to platinum-based chemotherapy, with one extending its analysis to differentiate between high and low risks of disease survival and platinum treatment resistance [16], [17]. Additional studies targeted various classification objectives, such as differentiating HGSC from non-HGSC [18], classifying epithelial OC into type I or II [19], and identifying OC as recurrent or nonrecurrent [20]. The majority of these studies were single-center initiatives, with a sample size ranging from a minimum of 6 patients to a maximum of 758 patients. However, to the best of our knowledge, this is the first attempt to predict TR exploiting a DL-based model in OC.
TAP CT images with different manufacturers (GE Medical Systems, Siemens, Philips, Toshiba, Hitachi) of 650 patients with OC treated between 2016 and 2022 were retrospectively collected at the European Institute of Oncology (IEO) in Milan, Italy. In this study, the TAP CT contrast enhanced portal-venous phase acquired at the moment of diagnosis was considered. The CTs in our dataset were meticulously manually annotated by 5 expert gynecologist for the purpose of classification. This involved a thorough examination of elecronic medical records, resulting in the assignment of TR = 0 and TR = 1 labels. TR=0 means complete tumor resection with no residual tumor, and the TR=1 means no-complete tumor resection with residual tumor. In our dataset, clinicians annotated 446 cases with TR=0 and 204 cases with TR=1. During the development of the model, all data were fully anonymized to ensure the utmost privacy and data protection. The inclusion and exclusion criteria for the study are illustrated in Table 1.
We performed images preprocessing techniques in Python 3.11. The following steps were performed: and therefore requires segmentation in the preprocessing phase. To address the limitations associated with manual segmentation, including potential bias, time-intensive procedures, and the scarcity of annotated data, we implement automatic segmentation using TotalSegmentator [21].
TotalSegmentator is a DL segmentation model which automatically and robustly segments all major anatomical structures in body CT images. Each organ is associated with a label, which allowed to set the upper and lower bounds of the ROI as ischiopubic rami of the pelvis and left and right hemidiaphragm cupola, respectively. Afterwards, each image was cropped along the z-axis according to the interval selected. The selection of the ROI derived from the fact that, compared to other tumors, OC metastasis occurs most frequently in the omentum or peritoneum, reporting almost 70% of patients with OC presented peritoneal cavity metastasis at the time of diagnosis.
Step #2: Additional standard preprocessing. Images pixels intensity was normalized between 0 and 1. Images were resized from the original dimension of 512 x 512 x n, with n varying among patients, to 128 x 128 x 128, where 128 was the average n.
The aforementioned preprocessing steps are illustrated in Figure 2.
For the classification task of predicting the binary clinical outcome TR from 3D TAP CT images, we designed a 3D CNN model. The architecture of the model comprises two fundamental components: a CNN-based Features Extractor (FE) and a feed-forward fully connected classifier. The FE is composed by a sequence of 7 convolutional blocks, each consisting of the following layers: a convolutional layer which increases the number of input channels, followed by a Rectified Linear Unit (ReLU) activation layer, a 3D batch normalization layer, another convolutional layer, which preserves the number of input feature maps, and a max-pooling layer which halves the spatial dimen- sions of the input. The FE takes as input preprocessed 3D TAP CT images with spatial dimensions of 128 x 128 x 128 and number of channels of 1, and returns a final vector of 512 extracted features. The feature vector is the input of the classifier, designed as a sequential model with linear layers interleaved with ReLU activation functions and Dropout layers, each having a dropout probability of 0.3. The overall architecture is shown in Figure 3.
The model combines the 3D CNN's ability to extract useful features from input 3D TAP CT images with the classifier's power to discriminate to predict the correct class.
We designed and trained our model from scratch, for the binary target task of TR classification in patients with OC.
We split our dataset into a training and validation set, with respectively 457 and 153 patients, and we evaluate our results on an external cohort of 40 patients. We configured a batch size 𝐵 = 8, employing Binary Cross Entropy (BCE) as a loss function, formulated as:
where 𝑦 is the model output and 𝑦 ˆis the target variable. The learning rate was set to be 0.0001, with a 0.1 multiplication every 30 epochs. Optimization was performed using the Adam algorithm, and the maximum training epoch was set to 200. The entire training procedure was executed on a single NVIDIA A100 GPU with 40GB of memory.
In this study, we employed a 5-fold cross-validation method to assess the performance of different model on different dataset splits based on the accuracy on the validation sets. We then proceeded to retrain this model using the best hyperparameters as found by the crossvalidation. After retraining, we evaluated its performance on a separate test set to confirm its effectiveness and generalization capabilities.
In figure 4, we report the confusion matrix. We obtained an accuracy value of 0.
In this paper, we delved into the power of DL models for the classification of TR in patients with OC, utilizing 3D TAP CT scans. TR is a pivotal diagnostic factor influencing clinical treatment decisions, that would highly improve the management of OC patients if accurately predicted at diagnosis. Leveraging the capabilities of DL in the medical domain, we extend its use to address this challenge in OC. Our methodology involves employing a 3D CCN model for binary TR classification, aiming to aid clinical decision in OC care. Previous studies have already involved and introduced DL in the context OC, but to the best of our knowledge, this is the first attempt to harness the potential of DL for specifically predicting TR. Indeed, one of the noteworthy challenges in this attempt is the absence of radiological indexes to inform total surgical tumor resection decisions.
Our main contribution is to address this gap, introducing a DL model as a non-invasive preoperative tool to facilitate clinical decision making.
In conclusion, it is important to recognize the limitations of our study, notably the potential for enhanced model generalization and performance by expanding the patient cohort, and considering alternative neural network architectures, such as Vision Transformer based models. We should broaden the application of our DL approach to predict other key diagnostic factors in OC, such as platinum sensitivity, overall survival, and surgical complications. Finally, the integration of explainability techniques should be essential for interpreting the model decisions, fostering trust, and promoting wider clinical use.
The authors declare that they have no conflict of interest.
This work is part of the PNRR-MAD-2022-12376574 project Under-XAI: understanding ovarian cancer initiation and progression through explainable AI, being exempted from the ethical committee approval by the National Ministry of Health. Furthermore, the European Institute of Oncology has implemented a broad consent which allows to include in the study all the institute's patients, except those that refused explicitly to sign the informed consent.
Informed consent was obtained from all individual participants included in the study.