Method for identifying unconventional cells in pathological section

Abstract

The invention discloses a method for identifying unconventional cells in a pathological section, comprising: preprocessing an electronic scanning pathological section to obtain an effective discriminating region, inputting the effective discriminating region into a full convolution network to be pre-trained, then replacing the full convolution network head fine-tuning network with a fully connected layer so that the full convolution network has an ability to extract unconventional cell features so as to determine the locations of unconventional cells and classify the effective discriminatingregions effectively; in combination with the prediction result vote of multiple common classification networks, outputting stable classification results. The identification method of the invention canautomatically discriminate the probability of existence of unconventional cells in each 20* magnified field of view in the pathological section, and outputs the unconventional cells with a probability value over 0.5 as a recognition result, greatly reduce workload of the artificial screening pathological section, and can quickly and accurately screen out the unconventional cells.

Description

technical field [0001] The invention belongs to the field of medical imaging, in particular to a method for identifying unconventional cells in pathological slices. Background technique [0002] Unconventional cells (or cells with abnormal shapes) in traditional pathological slices are screened manually: under a microscope, professional pathologists move the slices, and then scan the entire slice with naked eyes to find whether there are unconventional cells in the entire slice. This work is tedious and time-consuming, and as the reading time increases, the error rate also increases. [0003] With the continuous development of technology, the identification of unconventional cells in pathological slides can be initially screened with the help of computers. [0004] Based on the Convolutional Neural Network (CNN) algorithm, VGGNet, ResNet, DenseNet and other network improvement structures are constantly updated and iterated for computer vision. On natural images, its accurac...

AI Technical Summary

Problems solved by technology

[0002] Unconventional cells (or cells with abnormal shapes) in traditional pathological slices are screened manually: under a microscope, professional pathologists move the slices, and then scan the entire slice with naked eyes to find whether there are unconventional cells in the entire slice. This work is tedious and time-consuming, and as the reading time increases, the error rate also increases

Images