This study followed the STARD reporting guidelines and the Helsinki Declaration. The institutional review boards of the Sun Yat-sen Memorial Hospital approved this retrospective study with identified data [Institutional Review Board (IRB) no. SYSEC-KY-KS-2020-072] and issued a waiver for written consent. This study was registered with ClinicalTrials.gov, number NCT04511481. Data generated or analyzed during the study are available from the corresponding author on request.

Ninety-three patients from the Sun Yat-sen Memorial Hospital from January 2011 to January 2020 were enrolled and randomly assigned (2:1) to the training (n = 62) and validation cohorts (n = 31) ( Figure 1 ). Inclusion criteria: (i) pathologically confirmed EOC; (ii) whole abdomen contrast-enhanced MRI within 2 weeks before the initial intervention (MRI detailed in Supplementary Methods S1 and Table S1 ); (iii) received cisplatin/carboplatin chemotherapy (≥4 cycles) in the initial treatment or switched chemotherapy scheme due to primary platinum-refractory; and (iv) received either primary debulking surgery (PDS) or interval debulking surgery (IDS) (Residual disease <1 cm). Exclusion criteria: (i) cancer lesions invisible on MRI; (ii) poor quality MRI (e.g., with motion or artifacts) as assessed by two radiologists (ZW and ZC); (iii) concurrent cancers; and (iv) unavailability of complete clinical data.

Tumor response to chemotherapy was assessed by the RECIST 1.1 criteria before PDS; patients were followed up according to the NCCN guideline. The date of the last follow-up was 1 June 2021.

The primary endpoint was the platinum-free interval (PFI). According to the PFI, patients with ovarian cancer were divided into platinum-resistant (PFI <6 months) and platinum-sensitive (PFI ≥6 months). The secondary endpoint was progression-free survival (PFS), calculated from the first date of the initial intervention to disease progression or to the last progression-free date for those censored. Recurrences and distant metastases were confirmed by clinical, biochemical, or radiographic evidence or by biopsy findings.

We built prediction models based on different algorithms, MRI sequences, and volumes of interest (VOIs). The two representative deep learning models (whole abdomen model and primary tumor model) were constructed and validated with the same training and validation cohorts. The difference between the two models was the VOI in the preprocessing part ( Figure 2A ).

The whole abdomen model took the entire abdomen volume on axial CE-T1WI sequence (venous phase) and axial T2WI (acquired before the injection of contrast medium) as the VOI to be the input of the model without any segmentation and delineation by hand, which is the most significant difference to the primary tumor model. The, the sequences of an individual patient were first aligned into the same coordinate system in reference to the CE-T1WI sequence. Then, pixels outside of the human body area, which was defined by OTSU thresholding, were filled with Gaussian noise as in the original implementation of the Med3D network (20, 21).

The primary tumor model took the primary tumor area as the VOI on axial CE-T1WI and T2WI sequences. First, sequences were loaded into 3D Slicer software (version 4.10.2) in NRRD format and were then manually segmented and delineated slice-by-slice by two radiologists (ZW and ZC) independently as precisely as possible. The radiologists were blinded to prognostic information. The data was then reviewed by two radiologists (MG and ZW) from the same hospital, each of whom had more than 15 years of experience in gynecologic oncology radiology and was mainly responsible for the VOI assessment. Any disagreements were resolved by consensus among the four radiologists.

To eliminate the interference, all values from the areas outside of the VOIs were removed and filled with Gaussian noise, similar to the whole abdomen model ( Figure 2A ). Finally, spatial and intensity normalization was applied to these VOIs (see Supplementary Methods S2 ).

Features were extracted from the input data using a feed-forward neural network. It was pre-trained on a few medical imaging datasets to learn to capture useful information from medical images. Through the power of transfer learning, we are able to transfer the knowledge that the model has acquired from one task and apply it to the other. This means that the model does not need to learn everything from scratch every time. The neural network uses 3D convolution operations as its core structure. Hierarchically, deeper layers will produce more complicated features. For example while shallow layers may capture edges or brightness information, deeper layers may capture texture or even abnormality information. The model uses a residual structure, so that information extracted by the shallow layers will not be “forgotten” during the feeding forward through so many layers.

A pre-trained convolutional neural network Med3D (Resnet 10 version) was transferred as the feature encoder to extract features ( Figure 2B and Supplementary Methods S3 ) from the chosen VOI. Following the feature extraction on corresponding VOIs, we applied principal component analysis (PCA) to perform the decomposition ( Figure 2C and Supplementary Methods S4 ). It denoised and compressed them into smaller dimensions. Finally, a support vector machine (SVM) was fitted to the data produced by the PCA to achieve the classification of how likely a patient is to be platinum-sensitive ( Figure 2D and Supplementary Methods S5 ).

The validation cohort’s data was left completely untouched throughout the training process. In the building of the model pipeline, cross-validation was used to evaluate the model performance and to decide on the hyperparameters. All programs were run on Python version 3.6.8.

Feature maps were acquired from the output of the last convolution layer right before global average pooling. All the feature maps were averaged to get a heat map to determine which areas of the image were the most active and deserved the model’s focus ( Figure 2E ).

When assessing clinical characteristics, categorical variables were compared using the χ² test, and continuous variables were compared using the independent t-test. Receiver operating characteristic curve (ROC) analysis was used to determine the optimal threshold for maximizing prediction accuracy and to assess and evaluate the performance of models. Five-fold cross-validation was employed to compare robustness. The PFS was calculated using the Kaplan–Meier method, and the log-rank test, hazard ratios (HRs), and 95% confidence intervals (Cls) were calculated using a Cox regression analysis. The optimal cutoff values were generated with the R package ggsurvimier. Moreover, the calibration curve and decision curve analysis (DCA) were plotted to assess the whole abdomen model. P-values of less than 0.05 were considered significant, and all tests were 2-tailed. Statistical analyses were conducted with Python 3.6.8, and R version 3.6.1.

A total of 93 patients with EOC from the Sun Yat-sen Memorial Hospital were eligible for this study ( Figure 1 ). Table 1 shows the clinicopathological characteristics of patients from the training cohort (n = 62) and the validation cohort (n = 31), 50 (81%) of the 62 patients in the training cohort and 24 (77%) of the 31 patients in the validation cohort were platinum-sensitive. Hyperthermic intraperitoneal chemotherapy (HIPEC) was administered to 24 (39%) of the patients in the training cohort and 12 (39%) of the patients in the validation cohort. The median follow-up was 21.0 months (interquartile range [IQR] 13.7–33.2) for patients in the training cohort and 23.2 months (IQR 15.2–43.3) for the validation cohort. The clinicopathological variables between the training cohort and the validation cohort had no statistically significant difference (P >.05). In the univariate logistic regression analysis, stages were found to be associated with platinum sensitivity in the training cohort (P <.001) ( Table S2 ).

The performance of the primary tumor model compared with the whole abdomen model is summarized in Table S3 . The primary tumor model achieved areas under the receiver operating characteristic curves (AUCs) of 0.88 (95% CI, 0.79–0.97) and 0.81 (95% CI, 0.65–0.97), with accuracy (ACC) of 0.87 and 0.81, the sensitivity of 90% and 88%, and specificity of 75% and 57% in the training and validation cohorts, respectively. Five-fold cross-validation showed the specificity and sensitivity had an obviously reduced result ( Table S3 and Supplementary Figure S1 ).

In comparison, the whole abdomen model achieved AUCs of 0.97 (95% CI, 0.93–1.00) and 0.98 (95% CI, 0.93–1.00), with ACC of 0.95 and 0.97, the sensitivity of 96% and 96%, and the specificity of 92% and 100% in the training and validation cohorts, respectively. The robustness of the whole abdomen model was verified by five-fold cross-validation, which showed that ACC, sensitivity, and specificity maintained satisfactory levels ( Figures 3A, B and Table S3 ).

The calibration curve in Figure 3C indicated that the whole abdomen model was close to the perfect model and did not systematically make under- or over-predictions of platinum sensitivity, as shown by the fact that the Hosmer–Lemeshow test yielded a non-significant statistic for this model (P = .88 and P = .81 in the training and validation cohorts, respectively). The decision curve in Figure 3D indicates that the use of the whole abdomen model may improve clinical outcomes.

To further assess the whole abdomen model, stratified analysis was conducted ( Supplementary Figures S2–S5 ). The whole abdomen model achieved AUCs in the subgroup of Hyperthermic intraperitoneal chemotherapy (HIPEC) (AUCs 1.00, 0.94), non-HIPEC (AUCs 0.94, 1.00), stages III–IV (AUCs 0.98, 1.00), serous carcinoma (AUCs 0.97, 1.00), other pathological types (AUCs 1.00, 0.95), IDS (AUCs 1.00, 1.00), PDS (AUCs 0.97, 0.96) in the training and validation cohorts, respectively. In summary, the whole abdomen model was able to maintain stable performance across different subgroups.

A strong association between the whole abdomen model decision function value and progression-free survival (PFS) was further demonstrated by Kaplan–Meier analysis ( Figures 3E, F ). According to the decision function value, optimal cutoff values (0.296 and −1) were generated to split patients into high- and low-risk groups in the training and validation cohorts. High-risk patients achieved a lower decision function value and shorter PFS compared with the low-risk group in the training cohort (HR 0.17; 95% CI 0.08–0.36; P <.001, Log-rank test) and validation cohort (HR 0.12; 95% CI 0.03–0.41; P <.001, Log-rank test). To further substantiate the value of the whole abdomen model in prediction and prognosis, we used the model to predict 1- and 2-year PFS ( Figures 3G, H ). ROC curves indicated that in the 1-year PFS prediction, the AUCs were 0.96 and 0.98, and reduced to 0.72 and 0.71 in the 2-year PFS prediction in the training and validation cohorts, respectively.

As shown in Figure 4 , patients 1 and 2 were platinum-sensitive and belonged to the low-risk group that had higher decision function values, predictive probability, and longer PFS. In contrast, patients 3 and 4 were platinum-resistant and belonged to the high-risk group that had lower decision function value, predictive probability, and shorter PFS.

Although these high-level features were highly intricate, feature maps were transferred as heat maps and were observed and analyzed with the assistance of the deep learning visualization algorithms ( Figure 5 ). There were some high-response areas on the heat maps, which were red, yellow, and green. The red areas appeared to be more focused on the peritoneal space; the green areas seemed to be more focused on the obvious enhancement areas that demonstrated abundant vasculature (primary tumor, enlarged para-aortic lymph nodes, colon metastasis), and the yellow areas had a typical mesenteric shape ( Figure 5D ). These patterns seemed to be observed in most heat maps.

This research had various limitations. First, due to the limitation of sample size, this study lacked perspective and extensive data cohorts to ensure clinical applicability. Second, due to the retrospective nature, there was a lack of complete genetic data. Therefore, combining the genetic profiles and further explaining the deep learning features at the genetic level will become the focus of our future work.