TY - GEN
T1 - Joint Optimization of Convolutional Neural Network and Image Preprocessing Selection for Embryo Grade Prediction in In Vitro Fertilization
AU - Uchida, Kento
AU - Saito, Shota
AU - Pamungkasari, Panca Dewi
AU - Kawai, Yusei
AU - Hanoum, Ita Fauzia
AU - Juwono, Filbert H.
AU - Shirakawa, Shinichi
N1 - Publisher Copyright:
© 2019, Springer Nature Switzerland AG.
PY - 2019
Y1 - 2019
N2 - The convolutional neural network (CNN) is a standard tool for image recognition. To improve the performance of CNNs, it is important to design not only the network architecture but also the preprocessing of the input image. Extracting or enhancing the meaningful features of the input image in the preprocessing stage can help to improve the CNN performance. In this paper, we focus on the use of the well-known image processing filters, such as the edge extraction and denoising, and add the preprocessed images to the input of CNNs. As the optimal filter selection depends on dataset, we develop a joint optimization method of CNN and image processing filter selection. We represent the image processing filter selection by a binary vector and introduce the probability distribution of the vector. To derive the gradient-based optimization algorithm, we compute the gradients of weight and distribution parameters on the expected loss under the distribution. The proposed method is applied to an embryo grading task for in vitro fertilization, where the embryo grade is assigned based on the morphological criterion. The experimental result shows that the proposed method succeeds to reduce the test error by more than 8% compared with the naive CNN models.
AB - The convolutional neural network (CNN) is a standard tool for image recognition. To improve the performance of CNNs, it is important to design not only the network architecture but also the preprocessing of the input image. Extracting or enhancing the meaningful features of the input image in the preprocessing stage can help to improve the CNN performance. In this paper, we focus on the use of the well-known image processing filters, such as the edge extraction and denoising, and add the preprocessed images to the input of CNNs. As the optimal filter selection depends on dataset, we develop a joint optimization method of CNN and image processing filter selection. We represent the image processing filter selection by a binary vector and introduce the probability distribution of the vector. To derive the gradient-based optimization algorithm, we compute the gradients of weight and distribution parameters on the expected loss under the distribution. The proposed method is applied to an embryo grading task for in vitro fertilization, where the embryo grade is assigned based on the morphological criterion. The experimental result shows that the proposed method succeeds to reduce the test error by more than 8% compared with the naive CNN models.
KW - Convolutional neural network
KW - Embryo grading
KW - Image processing filter
KW - In vitro fertilization
UR - http://www.scopus.com/inward/record.url?scp=85076145662&partnerID=8YFLogxK
U2 - 10.1007/978-3-030-33723-0_2
DO - 10.1007/978-3-030-33723-0_2
M3 - Conference Proceeding
AN - SCOPUS:85076145662
SN - 9783030337223
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 14
EP - 24
BT - Advances in Visual Computing - 14th International Symposium on Visual Computing, ISVC 2019, Proceedings
A2 - Bebis, George
A2 - Parvin, Bahram
A2 - Boyle, Richard
A2 - Koracin, Darko
A2 - Ushizima, Daniela
A2 - Chai, Sek
A2 - Sueda, Shinjiro
A2 - Lin, Xin
A2 - Lu, Aidong
A2 - Thalmann, Daniel
A2 - Wang, Chaoli
A2 - Xu, Panpan
PB - Springer
T2 - 14th International Symposium on Visual Computing, ISVC 2019
Y2 - 7 October 2019 through 9 October 2019
ER -