Systematic Evaluation of Loss Functions for Ovarian Tumors Segmentation from Ultrasound Images

Ovarian cancer is one of the most dangerous gynecological diseases for women. Ovarian cancer is a malignant tumor that originates in one or both of ovaries. Today, the percentage of patients with ovarian cancer is increasing. Early detection of ovarian cancer can decrease the mortality rate. Therefore, recently, many approaches have concentrated on aiding in the diagnosis and detection of ovarian tumors from ultrasound images using deep learning models. In particular, the choice of a loss function of segmentation from ultrasound images is extremely important in training models. In this paper, we conduct a systematic evaluation of loss functions for ovarian tumor segmentation. To this end, we investigate the robustness of different loss functions including Binary Cross Entropy, Dice, Jaccard loss. A new loss named hybrid loss is introduced and integrated into a backbone framework based on U-Net and MobileNetV2. Experimental results on two benchmark datatset of ovarian tumor segmentation have confirmed that all these aforementioned loss functions perform well in balanced dataset. In case of imbalanced datasets, the proposed hybrid can be taken as a good choice.