Classification of histological types in ovarian cancer using deep learning.

BACKGROUND

Histological typing of carcinomas is crucial considering the varying progression, prognosis, and treatment efficacy. Clear cell and mucinous carcinomas show good early-stage prognosis but poor advanced-stage outcomes due to chemotherapy resistance. Here, we aimed to develop an artificial intelligence (AI) model to classify histological types of ovarian cancer.

METHODS

A deep learning model was developed using the YOLOv8 object detection (OD) and image classification. The OD and classification models were included in 71 cases of malignancy, comprising 2,438 and 5,874 images, respectively. In addition, 10 cases of 400 negative for malignancy (NFM) images were used for each model. The model was validated using test datasets that were not included in the model. The test dataset of the OD model included 37 cases comprising 957 images, and the test dataset of the classification model included 37 cases comprising 1,545 images of malignant cases. The NFM test dataset comprised 10 cases, each containing 200 images.

RESULTS

The OD model test dataset exhibited Precision, Recall, and F1 score values of 0.502, 0.482, and 0.492, respectively. The sensitivity of the model was 66.8%. The classification model test dataset showed sensitivities, specificities, and accuracies of 53.0%, 86.5%, and 79.2%, respectively. Of the 200 NFM images, malignant cells were misdetected in five images.

CONCLUSIONS

Accurate diagnosis of ovarian cancer using AI is challenging, as one histological type can mimic another, although the morphology that humans can diagnose can be classified with AI. Further studies are required to develop a high-accuracy AI model using three-dimensional imaging.