Room: AAPM ePoster Library
Purpose: Deep learning model observer (DLMO) has recently been enabled machine learning-based tool with emerging studies of artificial intelligence. In this paper, DLMO through a several convolution neural network layers of classification network have been used to compare a human observer performance. The purpose of this paper is to study DLMO on the DBT images of an anthropomorphic breast phantom and compare its performance to the real human observers using 4AFC study.
Methods: The classification network used in this study consists of five convolutional layers using convolution, pooling, and drop out operations and the dense layer with softmax activation function is followed for binary classification task. Two types of predicted output values were labeled by ‘0’ for lesion-absent and ‘1’ for lesion-present tasks. The images of spheroidal mass with different sizes were cropped by 200×200 size for their region of interest for 4AFC reading to the human observer and DLMO. A percentage of correct responses (PC) was measured at the end of each human observer test and compared by the accuracy of prediction using DLMO.
Results: The results indicated that our proposed DLMO showed the 93% accuracy using 60 testing datasets after 50 epochs of training using 204 input training datasets. Both the accuracy and loss function in training session of categorical cross-entropy reached a plateau after 35 epochs. The sensitivity and specificity of the testing data from DLMO showed 78% and 98%, respectively, which is comparable to the PC, 0.89, on 4AFC test from the human observer.
Conclusion: The proposed DLMO classification network maybe provide reasonable outcomes, compared to PC value from 4AFC human observer tests. However, the limited number of datasets used in this study has been remained a drawback. The future works will be directed to more quantitative analysis of DLMO using larger number of datasets.