Systems and methods of radiomics based cancer stratification
In the disclosed systems and methods for characterizing a cancer condition of a tissue in a subject, a computer system inputs information into an ensemble model. The information includes, for each respective class of radiomics features in a plurality of classes of radiomics features, a corresponding value for each respective radiomic feature in a corresponding plurality of radiomics features of the respective class of radiomics features obtained from a medical imaging dataset. The ensemble model comprises a plurality of component models. The computer system obtains as output from each respective component model in the plurality of component models a corresponding component prediction for the cancer condition, thereby obtaining a plurality of component predictions for the cancer condition. The computer system combines the plurality of component predictions to obtain as output of the ensemble model a characterization of the cancer condition.
1 . A method for characterizing a cancer condition of a tissue in a subject, the method comprising:
at a computer system that includes one or more processors and memory:
A) inputting information into an ensemble model comprising a plurality of component models to obtain as output from each respective component model in the plurality of component models a corresponding component prediction for the cancer condition, thereby obtaining a plurality of component predictions for the cancer condition, wherein:
the information comprises, for each respective class of radiomics features in a plurality of classes of radiomics features, a corresponding value for each respective radiomic feature in a corresponding plurality of radiomics features of the respective class of radiomics features obtained from a medical imaging dataset, wherein the medical imaging dataset comprises a plurality of medical images of the tissue in the subject acquired at a first time using a first medical imaging modality,
the ensemble model applies a plurality of parameters to the information through a plurality of computations, and
the inputting includes (i) inputting the corresponding value for each respective radiomics feature in the corresponding plurality of radiomics features of a first respective class of radiomics features in the plurality of classes of radiomics features into a first respective component model in the plurality of component models and (ii) inputting the corresponding value for each respective radiomics feature in the corresponding plurality of radiomics features of a second respective class of radiomics features in the plurality of classes of radiomics features into a second respective component model in the plurality of component models, wherein the corresponding value for no respective radiomics feature in the corresponding plurality of radiomics features of the first respective class of radiomics features is input into the second component model and the corresponding value for no respective radiomics feature in the corresponding plurality of radiomics features of the second respective class of radiomics features is input into the first respective component model; and
B) combining the plurality of component predictions to obtain as output of the ensemble model a characterization of the cancer condition.
2 . The method of claim 1 , wherein the characterization of the cancer condition comprises:
a respective cancer type selected from a plurality of cancer types,
a respective cancer stage selected from a plurality of cancer stages,
a respective tissue of origin selected from a plurality of tissues of origin,
a respective cancer grade selected from a plurality of cancer grades, or
a respective prognosis selected from a plurality of prognoses.
3 . The method of claim 1 , wherein the cancer condition is a cancer selected from the group consisting of a carcinoma, lymphoma, blastoma, glioblastoma, sarcoma, leukemia, breast cancer, squamous cell cancer, lung cancer, small-cell lung cancer, non-small cell lung cancer (NSCLC), adenocarcinoma of the lung, squamous carcinoma of the lung, head and neck cancer, cancer of the peritoneum, hepatocellular cancer, gastric or stomach cancer, pancreatic cancer, ovarian cancer, cervical cancer, liver cancer, bladder cancer, hepatoma, colon cancer, colorectal cancer, endometrial or uterine carcinoma, salivary gland carcinoma, kidney or renal cancer, liver cancer, prostate cancer, vulval cancer, thyroid cancer, hepatic carcinoma, B-cell lymphoma, low grade/follicular non-Hodgkin's lymphoma (NHL), small lymphocytic (SL) NHL, intermediate grade/follicular NHL, intermediate grade diffuse NHL, high grade immunoblastic NHL, high grade lymphoblastic NHL, high grade small non-cleaved cell NHL, bulky disease NHL, mantle cell lymphoma, AIDS-related lymphoma, Waldenstrom's Macroglobulinemia, chronic lymphocytic leukemia (CLL), acute lymphoblastic leukemia (ALL), hairy cell leukemia, and chronic myeloblastic leukemia.
4 . The method of claim 1 , wherein the plurality of component models is at least 40 component models.
5 . The method of claim 1 , wherein a respective component model in the plurality of component models is a neural network, a support vector machine, a Naive Bayes model, a nearest neighbor model, a boosted trees model, a random forests model, or a clustering model.
6 . The method of claim 1 , wherein each respective component model in the plurality of component models comprises a respective logistic regression model.
7 . The method of claim 1 , wherein the plurality of classes of radiomics features comprises:
a first subset of classes of radiomics features extracted from an unfiltered version of the plurality of medical images in the medical imaging dataset; and
a second subset of classes of radiomics features extracted from a filtered version of the plurality of medical images in the medical imaging dataset filtered by a first filtering methodology.
8 . The method of claim 7 , wherein the first filtering methodology comprises an imaging filter selected from the group consisting of a wavelet transform filter, a Laplacian of Gaussian (LoG) filter, a square transform filter, a square root transform filter, a logarithm transform filter, an exponential transform filter, a gradient transform filter, a 2-dimensional local binary pattern filter, and a 3-dimensional local binary pattern filter.
9 . The method of claim 7 , wherein the first subset of classes of radiomics features comprises at least 5 classes of radiomics features.
10 . The method of claim 9 , wherein the first subset of classes of radiomics features comprises a class of radiomics features selected from the group consisting of shape features, first order features, gray level cooccurrence matrix (GLCM) features, gray level run length matrix (GLRLM) features, gray level size zone matrix (GLSZM), gray level dependence matrix (GLDM) features, and neighboring gray tone difference matrix (NGTDM) features.
11 . The method of claim 9 , wherein the first subset of classes of radiomics features comprises shape features, first order features, gray level cooccurrence matrix (GLCM) features, gray level run length matrix (GLRLM) features, gray level size zone matrix (GLSZM), and gray level dependence matrix (GLDM) features.
12 . The method of claim 9 , wherein each respective class of radiomics features in the first subset of classes of radiomics features comprises at least 10, corresponding radiomics features.
13 . The method of claim 7 , further comprising, prior to the A) inputting the information into the ensemble model: 1) extracting, for each respective class of radiomics features in the first subset of classes of radiomics features, the corresponding value for each respective radiomic feature in the respective class of radiomics features in the first subset of classes of radiomics features from a region of interest (ROI) or a volume of interest (VOI) in the unfiltered version of the plurality of medical images; and 2) extracting, for each respective class of radiomics features in the second subset of classes of radiomics features, the corresponding value for each respective radiomic feature in the respective class of radiomics features in the second subset of classes of radiomics features from the ROI or the VOI in the filtered version of the plurality of medical images.
14 . The method of claim 13 , further comprising identifying the ROI or the VOI in the plurality of medical images, wherein identifying the ROI or the VOI in the plurality of medical images comprises:
segmenting the unfiltered version of the plurality of medical images into a plurality of segments or a plurality of volumes;
assigning, to each respective segment in the plurality of segments or to each respective volume in the plurality of volumes, a respective tissue classification in a plurality of tissue classifications based on one or more features of the respective segment or respective volume; and
grouping respective segments in the plurality of segments or respective volumes in the plurality of volumes assigned a target tissue classification in the plurality of tissue classifications, thereby identifying the ROI or the VOI.
15 . The method of claim 1 , wherein the medical imaging dataset comprises a computerized tomography (CT) dataset, a magnetic resonance imaging (MRI) dataset, an ultrasound dataset, a position emission tomography (PET) dataset, or an X-ray dataset.
16 . The method of claim 1 , wherein:
the characterization of the cancer condition is a respective characterization selected from a plurality of discrete characterizations of the cancer condition;
each respective component prediction in the plurality of component predictions is a respective component characterization of the cancer condition selected from the plurality of discrete characterizations of the cancer condition; and
the B) combining comprises identifying the respective component characterization that is most represented in the plurality of component predictions, thereby obtaining the characterization of the cancer condition.
17 . The method of claim 1 , wherein:
the characterization of the cancer condition is a respective characterization identified from a continuous range of characterizations of the cancer condition;
each respective component prediction in the plurality of component predictions is a respective component characterization of the cancer condition identified from the continuous range of characterizations; and
the B) combining comprises determining a measure of central tendency for the plurality of component predictions, thereby obtaining the characterization of the cancer condition.
18 . The method of claim 1 , wherein the B) combining comprises inputting the plurality of component predictions for the cancer condition into an aggregation model to obtain as output of the aggregation model the characterization of the cancer condition, wherein the aggregation model is a voting model with a tuned threshold.
19 . A computer system, comprising:
one or more processors; and
a non-transitory computer-readable medium including computer-executable instructions that, when executed by the one or more processors, cause the processors to perform a method comprising:
A) inputting information into an ensemble model comprising a plurality of component models to obtain as output from each respective component model in the plurality of component models a corresponding component prediction for the cancer condition, thereby obtaining a plurality of component predictions for the cancer condition, wherein:
the information comprises, for each respective class of radiomics features in a plurality of classes of radiomics features, a corresponding value for each respective radiomic feature in a corresponding plurality of radiomics features of the respective class of radiomics features obtained from a medical imaging dataset, wherein the medical imaging dataset comprises a plurality of medical images of the tissue in the subject acquired at a first time using a first medical imaging modality,
the ensemble model applies a plurality of parameters to the information through a plurality of computations, and
the inputting includes (i) inputting the corresponding value for each respective radiomics feature in the corresponding plurality of radiomics features of a first respective class of radiomics features in the plurality of classes of radiomics features into a first respective component model in the plurality of component models and (ii) inputting the corresponding value for each respective radiomics feature in the corresponding plurality of radiomics features of a second respective class of radiomics features in the plurality of classes of radiomics features into a second respective component model in the plurality of component models, wherein the corresponding value for no respective radiomics feature in the corresponding plurality of radiomics features of the first respective class of radiomics features is input into the second component model and the corresponding value for no respective radiomics feature in the corresponding plurality of radiomics features of the second respective class of radiomics features is input into the first respective component model; and
B) combining the plurality of component predictions to obtain as output of the ensemble model a characterization of the cancer condition.
20 . A non-transitory computer-readable storage medium having stored thereon program code instructions that, when executed by a processor, cause the processor to perform a method comprising:
A) inputting information into an ensemble model comprising a plurality of component models to obtain as output from each respective component model in the plurality of component models a corresponding component prediction for the cancer condition, thereby obtaining a plurality of component predictions for the cancer condition, wherein:
the information comprises, for each respective class of radiomics features in a plurality of classes of radiomics features, a corresponding value for each respective radiomic feature in a corresponding plurality of radiomics features of the respective class of radiomics features obtained from a medical imaging dataset, wherein the medical imaging dataset comprises a plurality of medical images of the tissue in the subject acquired at a first time using a first medical imaging modality,
the ensemble model applies a plurality of parameters to the information through a plurality of computations, and
the inputting includes (i) inputting the corresponding value for each respective radiomics feature in the corresponding plurality of radiomics features of a first respective class of radiomics features in the plurality of classes of radiomics features into a first respective component model in the plurality of component models and (ii) inputting the corresponding value for each respective radiomics feature in the corresponding plurality of radiomics features of a second respective class of radiomics features in the plurality of classes of radiomics features into a second respective component model in the plurality of component models, wherein the corresponding value for no respective radiomics feature in the corresponding plurality of radiomics features of the first respective class of radiomics features is input into the second component model and the corresponding value for no respective radiomics feature in the corresponding plurality of radiomics features of the second respective class of radiomics features is input into the first respective component model; and
B) combining the plurality of component predictions to obtain as output of the ensemble model a characterization of the cancer condition.