Methods for identifying biological material by microscopy
The present invention relates generally to the field of computer-based image recognition. More particularly, the invention relates to methods and systems for the identification, and optionally the quantitation of, discrete objects of biological origin such as cells, cytoplasmic structures, parasites, parasite ova, and the like which are typically the subject of microscopic analysis. The invention may be embodied in the form of a method for training a computer to identify a target biological material in a sample. The method may include accessing a plurality of training images, the training images being obtained by light microscopy of one or more samples containing a target biological material and optionally a non-target biological material. The training images are cropped by a human or a computer to produce cropped images, each of which shows predominantly the target biological material. A human then identifies the target biological material in each of the cropped images where identification is possible, and associating an identification label with each of the cropped images where identification was possible. A computer-implemented feature extraction method is then applied to each labelled cropped image. A computer-implemented learning method is then applied to each labelled cropped image to associate extracted features of abiological material with a target biological material.
1. A computer-implemented method for training a computer-implemented learning method to identify a genus or species of a parasite ovum in a sample, the method comprising:
accessing a plurality of computer-readable training images, the training images being obtained by light microscopy of one or more samples containing a parasite ovum and a non-parasite ovum material;
using a first subset of the plurality of computer-readable training images to perform human-supervised machine learning to distinguish images containing a parasite ovum from non-parasite ovum material to provide a machine learning model configured to distinguish parasite ovum material from non-parasite ovum material;
using a second subset of the plurality of computer-readable training images to perform frame cropping, the frame cropping comprising identifying a parasite ovum and cropping one or more of the plurality of computer-readable training images so as to produce one or more cropped computer-readable images, each of the one or more cropped computer-readable images showing predominantly the parasite ovum;
by human means identifying the genus or species of the parasite ovum in each of the one or more cropped computer-readable images where identification is possible;
associating an identification label with each of the one or more cropped computer-readable images where identification was possible;
applying a computer-implemented deep learning network feature extraction method to each labelled cropped computer-readable image; and
applying the machine learning model to each cropped computer-readable image to determine if the cropped computer-readable image contains a parasite ovum,
wherein the computer-implemented learning method is configured to associate one or more features extracted by the feature extraction method with a parasite ovum.
2. The computer-implemented method of claim 1 , wherein the cropping of one or more of the plurality of computer-readable training images comprises distinguishing the parasite ovum from the microscopy field background and/or the non-parasite ovum material, and cropping the parasite ovum such that the image comprises predominantly or substantially only parasite ovum.
3. The computer-implemented method of claim 2 , wherein the 1 distinguishing of the parasite ovum from the non-parasite ovum material is implemented at least in part by computer means.
4. The computer-implemented method of claim 3 , wherein the distinguishing of the parasite ovum from the non-parasite ovum material is by human-assisted computer means.
5. The computer-implemented method of claim 4 , wherein the human-assisted computer means comprises:
dividing a computer-readable training image into a series of frames,
by human means identifying the presence or absence of a parasite ovum in each of the series of frames,
applying a computer-implemented feature extraction method to each of the series of frames, and
applying a computer-implemented learning method to each of the series of frames,
wherein the computer-implemented learning method is configured to associate one or more extracted features with the presence or absence of a parasite ovum in a frame.
6. The computer-implemented method of claim 5 , wherein the association is used to identify the presence or absence of a parasite ovum in a plurality of frames that were not used in the human-assisted computer means, and where the parasite ovum is present cropping around the parasite ovum such that the image comprises predominantly or substantially only a parasite ovum.
7. The computer-implemented method of claim 1 , wherein each of the cropped images and/or features extracted therefrom is/are stored in a computer-readable database in linked association with its respective identification label.
8. The method of claim 1 , wherein the parasite ovum is the ovum of an intestinal parasite of a non-human animal.
9. The method of claim 1 , wherein the parasite ovum is selected from Haemonchus, Moniezia, Nematodirus, Ostertagia and Trichostrongulus.
10. The method of claim 1 that is capable of training a computer to correctly identify parasite ova to the following levels of precision: ≥98.5% for Haemonchus, 100% for Moniezia , and 100% for Nematodirus.