A prospective observational pilot study was carried out from November 2021to June 2023. Ethical clearance was obtained from the institutional ethics committee (no IECPG637/28.10.21). Patients were broadly classified in two groups: Group A, in which patients aged ≥ 18 years with suspected epithelial ovarian cancer were invited to participate. Exclusion criteria were non-epithelial ovarian cancer, patients receiving neoadjuvant chemotherapy, borderline ovarian tumour, incomplete radiological and pathological records, and those unwilling to participate.

The inclusion criteria for Group B were age ≥ 18 years, undergoing surgery for benign ovarian tumour, willing to participate in the study and no history of genital malignancy. The exclusion criteria for this group were patients undergoing conservative or medical management for benign ovarian tumour and those unwilling to participate in the study. Sociodemographic characteristics were recorded, clinical examination and pre operative investigations were done as per hospital protocol. All patients underwent one or more of the following radiological investigations, e.g., ultrasonography (USG), CT scan whole abdomen, magnetic resonance imaging (MRI) or PET-CT. Patients with radiological investigations done elsewhere were included in the study if they could provide the imaging records. The morphological features were evaluated by MRI or USG. USG findings were recorded by O-RADS reporting system (9). In case a patient had multiple radiological investigations, the radiologist selected the best images which were then uploaded in the CVAT software. Histopathology was considered as the gold standard for final recruitment as per the inclusion and exclusion criteria of the study. The study algorithm is summarized in Figure 1.

Datasets were obtained from the clinical features, radiological images and histopathological images of all 50 patients who underwent upfront surgery in Group A and Group B. Once the radiological images were uploaded, the images were manually annotated through computer vision annotation tool (CVAT) software. The collected dataset comprised 20 features used to diagnose ovarian tumours as either benign or malignant, of which 5 were baseline characteristics, i.e., age, menopausal status, duration of symptoms, parity, and ECOG performance score, while 12 were radiological features, i.e., size, characteristics, laterality, locularity, ascites, papillary projections, septations, omental disease, retroperitoneal nodes, peritoneal metastasis and O-RADS score and 3 blood biomarkers (serum CA125, CEA, CA19.9). Additionally, histology images of the malignant and benign cases were uploaded for algorithm evaluation. The features were selected on the basis of their relevance in detection and management of EOC (10, 11).

Data pre-processing was done to ensure compatibility with ML models and check for any inconsistencies. Recognizing the challenges of applying machine learning to a small oncology dataset, we adopted a cautious and clinically informed data splitting strategy. To test the developed models, the 50 patient dataset was split using the ‘train-test-split’ function. As the data was directly obtained from patients’ health records, pre-processing was done to handle inconsistencies and ensure compatibility with machine learning models. To address issues such as special symbols and incorrect values, we developed a class called remove special chars transformer. This class systematically identified and removed special symbols within the dataset, resulting in a clean dataset without such inconsistencies. There were no missing values recorded in the data set. To reduce the risk of overfitting, an important concern in small clinical cohorts, we deliberately restricted model complexity, applied regularization techniques, and monitored consistency across training and validation folds.

The voting selector algorithm (VSA) was used to improve the accuracy and reliability of feature subset identification by combining the outputs of multiple feature selection methods. The most suitable feature selection method was chosen by implementing and evaluating 11 different techniques. Out of these, three were filter methods, namely Mutual Information, ReliefF, and Monte Carlo; five were wrapper methods, namely, Forward Selection, Backward Selection, Boruta, Genetic Algorithm, and Simulated Annealing; lastly, three were embedded methods, namely, Recursive Feature Elimination, Ecological Model Rule Forest (EMRF), and Ecological Model Logistic Regression (EMLR). The VSA was applied and evaluated on Group A and Group B dataset to assess its performance in identifying informative feature subsets by the consensus voting process and maximize the model’s performance by selecting the hyperparameters. The fine tuning of the hyperparameter was performed using 5-fold cross-validation for four different algorithms: support vector machine (SVM), logistic regression (LR), random forest (RF) and artificial neural network (ANN). For the ANN algorithm, additional multiple layers of framework were tested. The best hyperparameters and their corresponding scores were then recorded. The GridSearchCV object performed a systematic search over this parameter grid using individual algorithm as the estimator. A total 3,000 plus image data sets were annotated including histology images of both benign and malignant ovarian tumours.

To test the developed models, the 50 patient dataset was split using the ‘train-test-split’ function.

Once the best hyperparameters for each model were obtained, the next step was to evaluate and compare the performance of the models on the classification task. Evaluation was based on the accuracy score, precision, recall and F1 score for measurement of functionality of each ML algorithm. The AUC (Area Under the Curve) measured the overall performance of the model across all thresholds. To derive a consensus feature selector algorithm, we combined the top four algorithms and considered the features selected by most of these algorithms. Thus the most effective model for the given classification task was identified. Finally, the ensemble learning method was used as a hard voting scheme with the top four models using the voting classifier function of sklearn library. Alongside traditional ML algorithms in a relatively limited dataset, additionally, we tested 2 deep learning algorithms ResNet50 (Residual Neural Network) and Inception Net v3 (GoogleNet) for bench marking and comparison with 4 ML algorithms. As the manuscript focused on the 4 ML algorithms, the details of the DL models were beyond the discussion in this manuscript.

After continuous epoch of all four ML algorithms, the best value was chosen for the each algorithm.

A total of 58 women were invited to participate in the study: Group A comprised of 30 women with epithelial ovarian cancer, while Group B comprised of 28 women with benign ovarian tumours. Eight patients had to be excluded for the following reasons: final histology leiomyoma (n = 1) and borderline ovarian tumour (n = 2), CD images not available (n = 4), withdrew consent (n = 1). Thus, 50 patients were included in the final analysis, with 25 women in each group. The sociodemographic characteristics of women participated in the study are summarized in Table 1.

The mean age of presentation in Group A was 49.60 ± 12.21 years and in Group B was 36.56 ± 14.95 years. The majority of women in both the groups had a ECOG ACRIN PS of 1 (56% vs. 84% in Group A and B respectively). The median duration of symptoms was 56 days (range 21–224 days) in Group A and 120 days (range 30–365 days) in Group B, with the most common presenting symptoms being distension of abdomen (44%) and mass per abdomen (40%), respectively. Among the risk factors, age and menopausal status were statistically significant (p < 0.05), whereas hormone intake, body mass index and family history of cancer were not statistically significant. Table 2 shows the radiological characteristics in both groups of patients.

Intraoperatively, the mean PCI value of the cancer cases was 9.64 ± 7.14. In 92% of cases, complete cytoreduction (CC) 0–1 was achieved. In 2 (8%) women, >2.5 cm residual disease had to be left behind because of peritoneal carcinomatosis. Out of 25 carcinoma ovary cases, 52% (n = 13) patients were of HGSOC, 24% (n = 6) LGSOC, 16% (n = 4) were mucinous carcinoma, and rest were endometrioid and clear cell carcinoma. For the benign counterpart, 40% (n = 10) were serous cystadenoma, 28% (n = 7) were mucinous cystadenoma, 20% (n = 4) endometrioma and 12% (n = 3) mature cystic teratoma. There was one case of serous cystadeno-fibroma on final histology. Adjuvant treatment for ovarian cancer cases were given according to the stage of the disease and following institutional protocol.

Figures 2A,B show the heat map describing the correlation between different independent radiological variables using Euclidean distance and average clustering method.

As shown in Figure 2A, the radiological findings namely, locularity, peritoneal metastases, papillary projections, septations and ascites were highly significant (1.5) for Group A, i.e., malignant cases; while colour score, O-RADS, size, laterality, and characteristics of the tumour were crucial in Group B, i.e., benign cases. Retroperitoneal nodes were more significant in malignant cases whereas p value was not significant when measured in standard logistic regression. Figure 2B shows that when average linkage clustering was used, radiological characteristics, namely, septations, peritoneal metastases and ascites were highly significant and were co-dependent (1.5) for Group A while, O-RADS, size, laterality, and characteristics were correlated (−1) features in Group B cases.

Table 3 shows the various performance metrics of all 4 ML algorithms.

All four ML algorithms showed high level of accuracy with maximum AUROC of 0.92 in RF model. Based on the crucial characteristics, RF and SVM were best performing when compared to counterparts LR and ANN. Both RF and SVM had an accuracy of 85.87 and 83.05%, respectively. For the DL set, both ResNet50 and Inception Net v3 had highest AUROC 0f 0.99.

Figure 3 depicts the ROC performances for all the ML algorithms.

The ROC curve was plotted to detect the sensitivity and specificity of all 4 ML algorithms. In addition, XG boost and Naïve Bayes were included for benchmarking and comparison with RF, LR, ANN and SVM. Out of all 4 ML algorithms, RF outperformed the other 3 algorithms with an AUROC of 0.92.