Analysis of Hyperparameter tuned UNet++ Deep model for Delineation of Ultrasound Ovarian Tumors
Ramita Shantharam, Rohini Palanisamy · 2023
Delineation of ovarian tumours in ultrasound images is essential in categorizing the incidence of ovarian cancer. In this paper, deep learning based hyperparameter tuned UNet++ model is utilized to segment abdominal ultrasound tumor images without human intervention. For this the 2D ultrasound images obtained from a public database is considered. These images are given as input to the UNet++ model. The hyperparameters of the model such as loss function, activation function and learning rate are analysed for their efficiency in precise segmentation of ovarian tumors. The model performance with the tuned hyperparameters was evaluated using dice coefficient and accuracy. Results indicate that the combination loss that combines Dice and Binary Cross Entropy (BCE) perform better than focal, dice and BCE individually by exhibiting a shorter interquartile range. The learning rate 0.0001 exhibited improved performance. The sigmoid activation function effectively introduced non-linearity when compared with ReLU and tanh activation functions. The resultant dice coefficient obtained was 84.6%. Thus, the proposed framework can be utilised to delineate ovarian tumors to assist clinical frameworks in automation and analysis.