Most studies of artificial intelligence in the medical field involve classification problems, but few consider recognition of one characteristic point in images or regression analysis such as data recognition. In this research, we constructed a fundamental convolutional neural network framework for regression analysis. Images of the handwritten digit "3" from the MNIST dataset were used as training data, with the protruding middle point as an image feature point. Input images and training data (x1, y1) were connected to 6 convolutional layers and then run through 2 affine layers to produce the output data (x2, y2). The loss function was the mean radial error (MRE) between the training and output data. After machine learning, the error converged to 0.75 pixels on average. We expect that this algorithm can be clinically applied to points having certain characteristics in images, such as locating hard tissue lesions or recognizing measurement points in cephalograms.

No abstract