Purpose
For lateral imaging of the knee joint, knowledge that the joint surfaces of the medial and lateral condyles are consistently drawn assumes diagnostic significance. If the difference between the joint surface of the medial and lateral condyles is within 5 mm, it is useful for diagnostic imaging.[1] However, measurement of the difference between the joint surfaces from the captured radiographic image requires the factor of measurement time because a human observer intervenes and performs the measurement.
Therefore, using the convolutional neural network (CNN),[2] which has...
Methods and materials
In this study, we built a network of CNN using Keras: library framework for NN. Table 1 shows the CNN architecture. The input tensor of the CNN is composed of (image_height, image_width, and image_channels). However, it does not include the batch dimensions. The output of each Conv2D layer and MaxPooling2D layer is a 3D tensor, with height, width, and channels as the shape. The width and height dimensions tend to shrink as the network gets deeper. The channel number is controlled by the first argument...
Results
The correct answer rate obtained with the evaluation data set using 1030 images was 82.4%, with an epoch of 2000, and L2 regularization of 0.1. Further details are shown in Fig. 3; Accuracy increased with increase in epoch. Test accuracy in Table 2 indicates the percentage of correct answers in the evaluation data set. The test was repeated five times with the same data set and learning conditions.
Fig. 4-6 shows the evaluation loss because of changing the L2 regularization parameters to 0.1, 0.01, and...
Conclusion
in this study, we report the behavior of a general-purpose NN library framework when limited clinical data was used for training. The CNN yielded a correct answer rate of 82.4% after analysis of only 1030 images. In addition, as shown in Fig. 4-6, it was found that a system with high generalization performance could be constructed with a small amount of training data. In future, if additional acquisition of clinical images and expansion of the training data set can facilitate the development of a more...
Personal information and conflict of interest
T. Ishibashi; Tokyo/JP - nothing to disclose M. Ogawa; Tokyo/JP - nothing to disclose A. Matsushima; Tokyo/JP - nothing to disclose N. Itayama; Tokyo/JP - nothing to disclose T. Okamoto; Tokyo/JP - nothing to disclose
References
Yu Yasuda, Hisaya Sato, Miwa Ohsawa et al. (2013) Proposal for an Auxiliary Tool Designed to Reduce Retake Rates for Lateral Radiography of the Knee Joint. JSRT. 69(10):1140-1145
Yoon Kim. (2019) Convolutional Neural Networks for Sentence Classification. arXiv:1408.5882.
Abien Fred Agarap. (2019) Deep Learning using Rectified Linear Units (ReLU). arXiv:1803.08375.
Sergey Ioffe, Christian Szegedy. (2015) Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift. arXiv:1502.03167.
Scott Cost, Steven Salzberg. (1993) A weighted nearest neighbor algorithm for learning with symbolic features. Machine Learning. Volume...