Cellular image analysis method, cellular image analysis device, and learning model creation method

Abstract

A phase image is formed by calculation from a hologram image of a cell, and segmentation is performed for each pixel for the phase image using a fully convolution neural network to identify an undifferentiated cell region, a deviated cell region, a foreign substance region, and the like. When learning, when a learning image included in a mini-batch is read, the image is randomly inverted vertically or horizontally and then is rotated by a random angle. A part that has been lost within the frame by the pre-rotation image is compensated for by a mirror-image inversion with an edge of a post-rotation image as an axis thereof. Learning of a fully convolution neural network is performed using the generated learning image. The same processing is repeated for all mini-batches, and the learning is repeated by a predetermined number of times while shuffling the training data allocated to the mini-batch. The precision of the learning model is thus improved. In addition, since rotationally invariant characteristics can be learned, it is possible to identify cell colonies of various shapes with good precision.