Training of machine-learning algorithms for digital pathalogy using healthy-tissue whole slide images
Various examples of the disclosure pertain to using whole-slide images that depict healthy tissue for a training process for at least one machine-learning algorithm for digital pathology. For instance, an autoencoder neural network can be trained based on the healthy tissue.
1 . A computer-implemented method of facilitating a training process of at least one machine-learning algorithm for digital pathology, the computer-implemented method comprising:
obtaining multiple whole-slide images exclusively depicting healthy tissue;
providing, to the training process, the multiple whole-slide images with an indication that the multiple whole-slide images exclusively depict healthy tissue; and
batch processing the multiple whole-slide images in an annotation process to automatically assign ground-truth labels to the multiple whole-slide images that identify each one of the multiple whole-slide images as exclusively depicting healthy tissue, wherein
the at least one machine-learning algorithm includes a novelty-detection algorithm including an auto encoder neural network algorithm,
the auto encoder neural network algorithm is exclusively trained on the multiple whole-slide images, and
an entirety of each of the multiple whole-slide images is provided to the training process.
2 . The computer-implemented method of claim 1 ,
wherein the training process comprises the annotation process for determining ground-truth labels for a plurality of whole-slide images, the plurality of whole-slide images comprising the multiple whole-slide images and multiple further whole-slide images, the multiple further whole-slide images depicting healthy tissue and unhealthy tissue.
3 . The computer-implemented method of claim 2 ,
wherein the annotation process comprises an interactive active-learning user-interaction protocol comprising multiple iterations, each iteration comprising at least one user query to obtain at least one respective ground-truth label, wherein user queries of subsequent iterations depend on user queries of preceding iterations, and
the batch processing of the multiple whole-slide images is implemented in a selected one of the multiple iterations.
4 . The computer-implemented method of claim 1 , further comprising:
executing the training process based on the multiple whole-slide images; and
upon training the novelty-detection algorithm, detecting unhealthy tissue by processing a further whole-slide image using the novelty-detection algorithm, unhealthy tissue being associated with abnormalities identified by the novelty-detection algorithm.
5 . The computer-implemented method of claim 4 ,
wherein the auto encoder neural network algorithm includes an encoder branch configured to encode an input data and a decoder branch configured to decode an encoded representation of the input data encoded by the encoder branch, and
the detecting detects the abnormalities based on a reconstruction error between the input data and an output data output by the decoder branch.
6 . The computer-implemented method of claim 1 ,
wherein the at least one machine-learning algorithm further comprises an anatomical modelling algorithm to rate an organ fitness of an organ anatomy in view of tumor growth, and the computer-implemented method further comprises:
executing the training process to train the anatomical modelling algorithm based on the multiple whole-slide images.
7 . The computer-implemented method of claim 6 , further comprising:
determining a cancer grade of a tumor based on a result of the novelty-detection algorithm and a further result of the anatomical modelling algorithm.
8 . The computer-implemented method of claim 1 , further comprising:
executing the training process in an unsupervised manner.
9 . The computer-implemented method of claim 1 , further comprising:
automatically determining one or more ground-truth labels based on the indication that the multiple whole-slide images exclusively depict healthy tissue; and
setting weights of the at least one machine-learning algorithm based on the multiple whole-slide images that depict healthy tissue and the one or more ground-truth labels.
10 . A non-transitory computer readable medium comprising program code, when executed by a processor, causes the processor to execute a method of facilitating a training process of at least one machine-learning algorithm for digital pathology, the method comprising:
obtaining multiple whole-slide images depicting healthy tissue;
providing, to the training process, the multiple whole-slide images with an indication that the multiple whole-slide images exclusively depict healthy tissue; and
batch processing the multiple whole-slide images in an annotation process to assign ground-truth labels to the multiple whole-slide images that identify each one of the multiple whole-slide images as exclusively depicting healthy tissue, wherein
the at least one machine-learning algorithm includes a novelty-detection algorithm including an autoencoder neural network algorithm,
the autoencoder neural network algorithm is exclusively trained on the multiple whole-slide images, and
an entirety of each of the multiple whole-slide images is provided to the training process.
11 . The non-transitory computer readable medium of claim 10 , wherein the program code, when executed by the processor, causes the processor to further execute:
automatically determining one or more ground-truth labels based on the indication that the multiple whole-slide images exclusively depict healthy tissue; and
setting weights of the at least one machine-learning algorithm based on the multiple whole-slide images that depict healthy tissue and the one or more ground-truth labels.
12 . The computer-implemented method of claim 3 , further comprising:
executing the training process based on the multiple whole-slide images; and
upon training the novelty-detection algorithm, detecting unhealthy tissue by processing a further whole-slide image using the novelty-detection algorithm, unhealthy tissue being associated with abnormalities identified by the novelty-detection algorithm.
13 . The computer-implemented method of claim 12 ,
wherein the auto encoder neural network algorithm includes an encoder branch configured to encode an input data and a decoder branch configured to decode an encoded representation of the input data encoded by the encoder branch, and
the detecting detects the abnormalities based on a reconstruction error between the input data and an output data output by the decoder branch.
14 . The computer-implemented method of claim 5 ,
wherein the at least one machine-learning algorithm further comprises an anatomical modelling algorithm to rate an organ fitness of an organ anatomy in view of tumor growth, and the computer-implemented method further comprises:
executing the training process to train the anatomical modelling algorithm based on the multiple whole-slide images.
15 . The computer-implemented method of claim 14 , further comprising:
determining a cancer grade of a tumor based on a result of the novelty-detection algorithm and a further result of the anatomical modelling algorithm.