Prediction of Eye Movement on Mammography with Convolutional LSTM

DOI
  • Okumura Eiichiro
    Department of Radiological Technology, Faculty of Health Sciences, Tsukuba International University
  • Kato Hideki
    Department of Radiological Science, Faculty of Health Science, Gunma Paz University
  • Honmoto Tsuyoshi
    Department of Radiological Technology, Ibaraki Children’s Hospital
  • Suzuki Nobutada
    Department of Radiology, Eastern Chiba Medical Center
  • Okumura Erika
    Department of Radiological Technology, Faculty of Health Sciences, Tsukuba International University
  • Higashigawa Takuji
    Group of Visual Measurement, Department of Technology, Nac Image Technology
  • Kitamura Shigemi
    Department of Radiological Technology, Faculty of Health Sciences, Tsukuba International University
  • Muranaka Hiroyuki
    Department of Radiological Technology, Faculty of Health Sciences, Tsukuba International University
  • Ando Jiro
    Mammary surgery, Tochigi Cancer Center
  • Ishida Takayuki
    Division of Health Sciences, Graduate School of Medicine, Osaka University

Bibliographic Information

Other Title
  • Convolutional LSTMを用いた乳房画像の視線動向の予測

Search this article

Description

<p>It is very difficult for radiologist to correctly detect small calcifications and lesions hidden in dense breast tissue on mammography. There are previous papers that detect lesions of the observer’s eye-tracking information in chest radiography etc. by CNN. Therefore, we investigated in 3ch convLSTM, Autoencoding convLSTM, and U-net convLSTM for deep learning, and aimed to predict the eye-tracking movement in mammography with high accuracy. We obtained gaze-tracking data for four mammography expert radiologists and 15 mammography technologists on 15 abnormal and 15 normal mammographies published by the MIAS. Next, a heat map was created at 4-second intervals, and 3ch convLSTM, Autoencoding convLSTM, and U-net convLSTM was used to predict the heat map image 4 seconds ahead from the temporal two heat map images. In the SSIM in U-net convLSTM, 4-8 seconds to 16-20 seconds was 0.96±0.01. In all 4-8 seconds to 16-20 seconds, the SSIM in U-net convLSTM was higher than this in 3ch convLSTM, Autoencoder convLSTM and there was a statistically significant difference (P<0.05). In the future, it will be necessary to increase the number of cases and further improve the prediction.</p>

Journal

Keywords

Details 詳細情報について

Report a problem

Back to top