Aims:
Artificial intelligence is rapidly gaining ground in online detection, endoscopic
and morphological characterization of colon epithelial neoplasms. Even for pathologists
identification of metaplasia and dysplasia in the epithelium of the mucous glands
could be an extremely difficult task. The same task in vivo, directly during the endoscopic
examination is no less difficult, therefore the development of auxiliary mathematical
models for image recognition is requested.
Methods:
We propose a new design of a convolutional neural network (CNN) based on U-Net model
and use it for mucous glands segmentation. The main distinctive ideas of the proposed
CNN lay in the multiscale architecture, using non-local blocks to capture long-range
dependencies in the image and using a contour-aware loss function. The network was
first trained on the public Warwick-QU dataset with non-linear augmentation process
and was afterthat fine-tuned on the manually labeled histological images obtained
from paraffin sections of endoscopic biopsy material of the colon.
Results:
The multiscale architecture of the proposed segmentational CNN makes it less sensitive
to the scale of the input image. Due to the specific loss function it is able to detect
and separate “stuck” glands. The used non-linear blocks have a positive effect on
the time needed for model to converge. Altogether this leads to the accurate segmentation
of glands on histology images (Dice coefficient = 0.87 for Warwick-QU dataset, Dice
coefficient = 0.83 for the obtained dataset).
Conclusions:
The generalization ability of the proposed algorithm enables it to effectively segment
individual glands as well as to perform inner-gland segmentation (detect nuclei, lumen
and cytoplasm) in histological images. The subsequent development of this gland segmentation
technology can allow to detect changes in the lumen shape (serration) of glands, in
the nuclear-cytoplasmic ratio inside mucus-forming cells, and in the character of
the expression of immunohistochemical markers.
