3D Supervised Contrastive-Learning Network for Classification of Ovarian Neoplasms

Tarun Kanti Roy, Suely Oliveira, Jesus Gonzalez Bosquet, Xiaodong Wu

Keywords: Supervised contrastive learning, ovarian cancer, classification, deep learning, feature encoder, densenet, resnet, cross validation, tumors, CT scan

TL;DR: We propose a deep learning based predictive model for ovarian cancer patients by implementing 3D contrastive learning.

Abstract:

Ovarian cancer ranks the $5^{th}$ in cancer deaths among women, accounting for more deaths than any other cancer of the female reproductive system.
We propose a 3D contrastive learning based predictive model to discriminate benign from malignant masses in abdominal CT scans for ovarian cancer patients. We used fully supervised contrastive learning(SCL) approach which allowed us to effectively leverage the label information of our small dataset of 331 patients.
All patients' data was collected at the University of Iowa.
Three different architectures (VGG, ResNet and DenseNet) were implemented for feature extraction by contrastive learning. We showed that SCL consistently out-performed over the traditional cross-entropy based networks with VGG and two ResNet variants. With five fold cross validation, our best contrastive learning model achieves an accuracy of 92.8\%, mean AUC of 92.4\%, mean recall of 94.45\% and mean specificity of 90.37\%. This work shows that contrastive learning is a promising deep learning method to improve early detection of women at risk of harboring ovarian cancer.