Ovarian cancer (OC) is the most lethal gynecologic malignancy worldwide, characterized by aggressive behavior, high relapse rate, and rapid progression. The cornerstone of OC treatment is cytoreductive surgery, targeting the removal of all detectable tumor lesions wherever feasible. In instances of widespread disease or significant perioperative morbidity risk, patients may initially receive neoadjuvant chemotherapy aimed at reducing the tumor's volume prior to surgical intervention. The pivotal decision between surgery and chemotherapy poses a significant therapeutic challenge in OC management. Our contribution is to develop an artificial intelligence-based model to support this critical decision by predicting Tumor Resectability (TR) from preoperative Computed Tomography (CT) images at the time of diagnosis. Our study aims to develop a 3D Convolutional Neural Network capable of predicting TR in a cohort of 650 with advanced stage epithelial patients with OC who underwent surgery at the European Institute of Oncology (IEO, Milan, Italy). The model processes preoperative CT scans of the Thorax, Abdomen, and Pelvis to deliver a binary prediction: TR=0 indicates a tumor completely resected, while TR=1 indicates the presence of residual tumor after cytoreductive surgery. We design and train our model from the ground up, achieving as preliminary results an accuracy of 65%. As far as we are aware, this is the first attempt to leverage deep learning for assessing TR in OC patients based on preoperative CT scans. Our model represents a non-invasive and preoperative tool with the potential to facilitate clinical decision making in the era of individualized and precision medicine. The work is part of the project Under-XAI: understanding ovarian cancer initiation and progression through explainable AI. Project code: PNRR-MAD-2022-12376574.
3,
this standard of care, it is estimated that 70% of patients
Ovarian Cancer (OC) is the most lethal gynaecologic ma- will experience a relapse. Surgical intervention aims at
lignancy worldwide, ranking as the fifth deadliest cancer achieving complete tumor resection; however, it often
among women and accounting for approximately 13000 results to be either aggressive, leading to severe
postopdeaths in 2023 in the United States [
Artificial Intelligence (AI) has been demonstrated to enhance the efectiveness 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 surFigure 1: In these CT scans, 4 diferent patients with OC are vival in 364 epithelial OC patients. In this study, 657 depicted, each presenting unique diagnostic challenges. The quantitative descriptors were extracted from preoperafirst patient’s scan identifies a discrete retroperitoneal lymph tive CT images, upon which the ML algorithm Radiomic node measuring 8.5 mm. The second patient has a conspicuous Prognostic Vector was developed. The latter accurately omental cake, which is considerably larger at 29.2 mm. For identified the 5% of patients with a median overall surthe third patient, a 13.3 mm nodule is present. Lastly, the vival of less than 2 years, demonstrating significant imfourth patient’s scan shows peritoneal thickening of 10.5 mm, provement over established prognostic methods. Crispinraising concerns about potential peritoneal carcinomatosis. Ortuzar et al. [12] addressed the challenge of predicting Each case demonstrates the diverse presentations of OC and neoadjuvant chemotherapy (NACT) response of 72 highthe inherent challenges in predicting TR preoperatively. grade serous OC (HGSOC) afected patients, presenting an ensemble ML model that, integrating baseline clinical, blood-based, and radiomic biomarkers from primary and part of personalized oncology treatments, driven by ad- metastatic lesions, predicted changes in total disease volvancements in Machine Learning (ML)[3]. However, ML ume. Validation on internal and external cohorts showed requires appropriate selection among the numerous ra- that the model significantly improved prediction accudiomic features extracted from images [4] [5]. racy compared to the clinical model, highlighting the Deep Learning (DL) has shown promising results in au- potential of radiomics in enhancing treatment response tomatically and directly detecting valuable features from predictions. medical imaging [6], boosting the progress of computer On the front of DL, Jan et al.[10] developed an AI ensemvision in the medical field [ 7] [8], and demonstrating su- ble model combing radiomics, DL and clinical features perior performance in comparison to hand-crafted image from CT images to distinguish between benign and mafeatures [9]. lignant OC. With 149 patients and 185 tumors, the model In this paper, a 3D CNN was designed to perform bi- achieved 82% accuracy, 89% specificity, and 68% sensitivnary TR classification of patients with OC. Specifically, ity. Compared to junior radiologists, the model exhibited what emerged from literature research was the absence higher accuracy and specificity while maintaining comof robust radiological indexes to select patients for total parable sensitivity. Wang et al.[7] proposed a DL method surgical resection. Hence, our primary contribution was to predict 3-year recurrence in 245 high-grade serous OC the implementation of a non-invasive and preoperative patients from preoperative CT images. The DL network, DL model to assess whether an upfront patient could be a trained on 8917 CT images, extracts a 16-dimensional suitable candidate for debulking surgery, when achieving DL feature used to predict the outcome probability. The a total resection appears feasible, or the patient might model achieved AUC values of 0.772 and 0.825 for high be recommended to undergo neoadjuvant chemotherapy and low recurrence risk, exhibiting stronger prognosbefore proceeding to interval surgery, when a complete tic value compared to clinical characteristics. Zheng et resection seems unlikely. Therefore, the proposed model al.[13] proposed a Vit-based DL model for predicting might potentially assist radiologists and gynecologists to overall survival in 734 high-grade serous OC patients assess TR and guide therapeutic strategies for patients using preoperative CT images. Analyzing 734 patients, with OC. 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 reso- Table 1 nance imaging (MRI). A pre-trained CNN were used and Inclusion and Exclusion Criteria 1,024 features were automatically extracted from MRI Inclusion Criteria sequences to predict platinum sensitivity.The model performed Area Under the Curve (AUC) of 0.97 and 0.98 in Epithelial OC training and validation cohorts. Advanced stage (III-IV) Among the 20 research papers examining OC in [15], 11 CT acquired before treatment primarily aimed at classifying between benign, malig- Age ≥ 18 years nant, and/or borderline tumors. Two of these studies focused on resistance to platinum-based chemotherapy, with one extending its analysis to diferentiate between high and low risks of disease survival and platinum treatment resistance [16], [17]. Additional studies targeted various classification objectives, such as diferentiating 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.
3.0.1. Dataset CT slice thickness > 5 mm
No consent to research
No CT or data available 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.
TAP CT images with diferent 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 2. Step #2: Additional standard preprocessing. Immoment of diagnosis was considered. The CTs in our ages pixels intensity was normalized between 0 dataset were meticulously manually annotated by 5 ex- and 1. Images were resized from the original dipert gynecologist for the purpose of classification. This mension of 512 x 512 x n, with n varying among involved a thorough examination of elecronic medical patients, to 128 x 128 x 128, where 128 was the records, resulting in the assignment of TR = 0 and TR = average n. 1 labels. TR=0 means complete tumor resection with no The aforementioned preprocessing steps are illustrated residual tumor, and the TR=1 means no-complete tumor in Figure 2. resection with residual tumor. In our dataset, clinicians annotated 446 cases with TR=0 and 204 cases with TR=1. 3.0.3. Model architecture During the development of the model, all data were fully anonymized to ensure the utmost privacy and data pro- For the classification task of predicting the binary clinical tection. outcome TR from 3D TAP CT images, we designed a 3D The inclusion and exclusion criteria for the study are CNN model. The architecture of the model comprises two illustrated in Table 1. fundamental components: a CNN-based Features Extractor (FE) and a feed-forward fully connected classifier. The 3.0.2. Image pre-processing FE is composed by a sequence of 7 convolutional blocks, each consisting of the following layers: a convolutional We performed images preprocessing techniques in layer which increases the number of input channels, folPython 3.11. The following steps were performed: lowed 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 dimen1. Step #1: Segmentation and Region of Interest (ROI) selection. Identifying the ROI most afected by OC in CT images is a fundamental first step, on diferent 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 efectiveness and generalization capabilities.
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 6. Conclusions 512 extracted features. The feature vector is the input of the classifier, designed as a sequential model with linear In this paper, we delved into the power of DL models layers interleaved with ReLU activation functions and for the classification of TR in patients with OC, utilizing Dropout layers, each having a dropout probability of 0.3. 3D TAP CT scans. TR is a pivotal diagnostic factor inThe overall architecture is shown in Figure 3. lfuencing clinical treatment decisions, that would highly The model combines the 3D CNN’s ability to extract improve the management of OC patients if accurately useful features from input 3D TAP CT images with the predicted at diagnosis. Leveraging the capabilities of DL classifier’s power to discriminate to predict the correct in the medical domain, we extend its use to address this class. challenge in OC. Our methodology involves employing We designed and trained our model from scratch, for the a 3D CCN model for binary TR classification, aiming to binary target task of TR classification in patients with aid clinical decision in OC care.
OC. 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 4. Experiment description specifically predicting TR. Indeed, one of the noteworthy challenges in this attempt is the absence of radiological We split our dataset into a training and validation set, indexes to inform total surgical tumor resection decisions. with respectively 457 and 153 patients, and we evaluate Our main contribution is to address this gap, introducing our results on an external cohort of 40 patients. a DL model as a non-invasive preoperative tool to faciliWe configured a batch size = 8, employing Binary tate clinical decision making.
Cross Entropy (BCE) as a loss function, formulated as: In conclusion, it is important to recognize the limita tions of our study, notably the potential for enhanced 1 ∑︁ [ · log(ˆ) + (1 − ) · log(1 − ˆm)]odel generalization and performance by expanding the BCE(ˆ, ) = − =1 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. where is the model output and ˆ is 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 diferent model
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.
patients with advanced ovarian cancer following for predicting overall survival in patients with highupfront radical debulking surgery, Gynecologic grade serous ovarian cancer, Frontiers in Oncology oncology 141 (2016) 264–270. 12 (2022) 986089. [3] B. J. Erickson, P. Korfiatis, Z. Akkus, T. L. Kline, Ma- [14] R. Lei, Y. Yu, Q. Li, Q. Yao, J. Wang, M. Gao, Z. Wu, chine learning for medical imaging, radiographics W. Ren, Y. Tan, B. Zhang, et al., Deep learning 37 (2017) 505–515. magnetic resonance imaging predicts platinum sen[4] R. J. Gillies, P. E. Kinahan, H. Hricak, Radiomics: sitivity in patients with epithelial ovarian cancer, images are more than pictures, they are data, Radi- Frontiers in Oncology 12 (2022) 895177. ology 278 (2016) 563–577. [15] P. Shrestha, B. Poudyal, S. Yadollahi, D. E. Wright, [5] M. R. Tomaszewski, R. J. Gillies, The biological A. V. Gregory, J. D. Warner, P. Korfiatis, I. C. Green, meaning of radiomic features, Radiology 298 (2021) S. L. Rassier, A. Mariani, et al., A systematic review 505–516. on the use of artificial intelligence in gynecologic [6] B. J. Erickson, P. Korfiatis, T. L. Kline, Z. Akkus, imaging–background, state of the art, and future K. Philbrick, A. D. Weston, Deep learning in radiol- directions, Gynecologic Oncology 166 (2022) 596– ogy: does one size fit all?, Journal of the American 605.
College of Radiology 15 (2018) 521–526. [16] H. Veeraraghavan, H. A. Vargas, A. Jimenez[7] S. Wang, Z. Liu, Y. Rong, B. Zhou, Y. Bai, W. Wei, Sanchez, M. Micco, E. Mema, Y. Lakhman, M. Wang, Y. Guo, J. Tian, Deep learning pro- M. Crispin-Ortuzar, E. P. Huang, D. A. Levine, vides a new computed tomography-based prognos- R. N. Grisham, et al., Integrated multi-tumor radiotic biomarker for recurrence prediction in high- genomic marker of outcomes in patients with high grade serous ovarian cancer, Radiotherapy and serous ovarian carcinoma, Cancers 12 (2020) 3403.
Oncology 132 (2019) 171–177. [17] X.-p. Yu, L. Wang, H.-y. Yu, Y.-w. Zou, C. Wang, J.[8] D. S. Kermany, M. Goldbaum, W. Cai, C. C. Valentim, w. Jiao, H. Hong, S. Zhang, Mdct-based radiomics H. Liang, S. L. Baxter, A. McKeown, G. Yang, X. Wu, features for the diferentiation of serous borderline F. Yan, et al., Identifying medical diagnoses and ovarian tumors and serous malignant ovarian tutreatable diseases by image-based deep learning, mors, Cancer Management and Research (2021) cell 172 (2018) 1122–1131. 329–336. [9] M. S. Choi, B. S. Choi, S. Y. Chung, N. Kim, J. Chun, [18] H. An, Y. Wang, E. M. Wong, S. Lyu, L. Han, J. A.
Y. B. Kim, J. S. Chang, J. S. Kim, Clinical evalu- Perucho, P. Cao, E. Y. Lee, Ct texture analysis in ation of atlas-and deep learning-based automatic histological classification of epithelial ovarian carsegmentation of multiple organs and clinical tar- cinoma, European Radiology 31 (2021) 5050–5058. get volumes for breast cancer, Radiotherapy and [19] L. Qian, J. Ren, A. Liu, Y. Gao, F. Hao, L. Zhao, H. Wu, Oncology 153 (2020) 139–145. G. Niu, Mr imaging of epithelial ovarian cancer: [10] Y.-T. Jan, P.-S. Tsai, W.-H. Huang, L.-Y. Chou, S.-C. a combined model to predict histologic subtypes, Huang, J.-Z. Wang, P.-H. Lu, D.-C. Lin, C.-S. Yen, European Radiology 30 (2020) 5815–5825. J.-P. Teng, et al., Machine learning combined with [20] Y. Liu, Y. Zhang, R. Cheng, S. Liu, F. Qu, X. Yin, radiomics and deep learning features extracted from Q. Wang, B. Xiao, Z. Ye, Radiomics analysis of ct images: a novel ai model to distinguish benign apparent difusion coeficient in cervical cancer: a from malignant ovarian tumors, Insights into Imag- preliminary study on histological grade evaluation, ing 14 (2023) 68. Journal of Magnetic Resonance Imaging 49 (2019) [11] H. Lu, M. Arshad, A. Thornton, G. Avesani, P. Cun- 280–290.
nea, E. Curry, F. Kanavati, J. Liang, K. Nixon, [21] J. Wasserthal, H.-C. Breit, M. T. Meyer, M. Pradella, S. T. Williams, et al., A mathematical-descriptor D. Hinck, A. W. Sauter, T. Heye, D. T. Boll, J. Cyrof tumor-mesoscopic-structure from computed- iac, S. Yang, et al., Totalsegmentator: Robust segtomography images annotates prognostic-and mentation of 104 anatomic structures in ct images, molecular-phenotypes of epithelial ovarian cancer, Radiology: Artificial Intelligence 5 (2023).
Nature communications 10 (2019) 764. [12] M. Crispin-Ortuzar, R. Woitek, M. A. Reinius,
E. Moore, L. Beer, V. Bura, L. Rundo, C. McCague, S. Ursprung, L. Escudero Sanchez, et al., Integrated radiogenomics models predict response to neoadjuvant chemotherapy in high grade serous ovarian cancer, Nature communications 14 (2023) 6756. [13] Y. Zheng, F. Wang, W. Zhang, Y. Li, B. Yang, X. Yang,
T. Dong, Preoperative ct-based deep learning model