Method and system for automatic classification of radiographic images having different acquisition characteristics
A method and system are disclosed for generating a machine learning model for automatic classification of radiographic images acquired by various acquisition protocols. The method includes the steps of: providing a plurality of radiographic images, detecting and segmenting in each of the radiographic image at least one regions of interest (ROI) as reference ROI, measuring at least one radiomic feature per reference ROI, identifying valid reference ROIs based on the measured radiomics values, and clustering the measured radiomics values of valid reference ROIs into at least two reference clusters according to a set of characteristics of image acquisition. A method and system are disclosed for classifying radiographic images by applying a machine learning model generated for automatic classification of radiographic images.
1 . A method for generating a machine learning model for automatic classification of radiographic images acquired by various acquisition protocols, comprising:
Providing a plurality of radiographic images of at least two different patients, the radiographic images of each of the at least two different patients being acquired with acquisition protocols different from each other,
Detecting and segmenting in each of said radiographic images at least one regions of interest (ROI) as reference ROI,
Measuring at least one radiomic feature per reference ROI,
Identifying valid reference ROIs from a plurality of reference ROIs based on the measured radiomics values,
Clustering the measured radiomics values of said valid reference ROIs into at least two reference clusters according to a set of characteristics of image acquisition,
wherein the plurality of radiographic images comprises at least one kind of the following: clinical images being acquired by imaging a patient, phantom images acquired by imaging a phantom object, and mixed images each combining a patient and a pocket phantom acquired by imaging a patient with a phantom object in proximity to the patient, and wherein said clinical images are used to generate a first machine learning model and said phantom images and/or said mixed images are used to generate a second machine learning model.
2 . The method according to claim 1 , further comprising, before detecting and segmenting at least one ROIs as reference ROI, pre-processing each of said radiographic images by using at least one of the following image processing techniques: denoising, resampling, diffusion-weighted magnetic resonance imaging and apparent diffusion coefficient mapping.
3 . The method according to claim 1 , wherein detecting and segmenting in each of said radiographic images at least one regions of interest as reference ROI comprises, if the radiographic images are clinical images, detecting and segmenting in each clinical images at least one non-pathological regions of interest as reference ROI.
4 . The method according to claim 3 , wherein detecting and segmenting in each clinical image at least one non-pathological regions of interest as reference ROI, comprises segmenting various non-pathological anatomical structure of interest (ASIs), and segmenting and detecting ROIs in the segmented non-pathological ASIs.
5 . The method according to claim 1 , wherein detecting and segmenting in each of said radiographic images at least one regions of interest as reference ROI comprises, if the radiographic images are phantom images and/or mixed images each combining a patient and a pocket phantom, segmenting various phantom-derived structure of interest (PDSI) based on a reference material, and segmenting and detecting ROIs in the segmented PDSIs.
6 . The method according to claim 1 , further comprises, after detecting and segmenting at least one non-pathological regions of interest as reference ROI, optimizing at least one of the previously detected and segmented ROIs by at least one image processing technique chosen among binning of image values and rescaling.
7 . The method according to claim 1 , wherein measuring at least one radiomic feature per reference ROI comprises radiomics computing of each reference ROI.
8 . The method according to claim 1 , wherein identifying valid reference ROIs from a plurality of reference ROIs further comprises comparing the measured radiomics values to a predetermined range of radiomics values chosen from a set of reference value ranges and determining a valid reference ROI if the measured radiomics values thereof fall into the predetermined range of radiomics values.
9 . The method according to claim 1 , wherein clustering measured radiomics values of said valid reference ROIs into at least two reference clusters comprises classifying the measurements of valid reference ROIs into at least two clusters, and wherein clustering measured radiomics values of said valid reference ROIs into at least two reference clusters further comprises characterizing the at least two clusters in order to obtain at least two reference clusters.
10 . A method for classifying at least one radiographic images according to a set of characteristics of image acquisition by applying a machine learning model generated for automatic classification of radiographic images, said machine learning model defining at least two reference clusters of said radiographic images acquired by various acquisition protocols, comprising:
Receiving at least one radiographic image,
Providing one of the at least one radiographic image as a candidate image for classification,
Detecting and segmenting in said candidate image at least one region of interest as candidate reference ROI,
Measuring at least one radiomic feature per candidate reference ROI,
Identifying valid candidate reference ROIs from a plurality of candidate reference ROIs based on the measured radiomics values,
Comparing the measured radiomics values of each said valid candidate reference ROIs to each of the radiomics value of each of said at least two reference clusters, and
Classifying the candidate image into one of the at least two reference clusters based on the comparison.
11 . The method according to claim 10 , wherein the plurality of radiographic images comprises at least one kind of the following: clinical images being acquired by imaging a patient, phantom images acquired by imaging a phantom object, and mixed images each comprising a patient and a pocket phantom acquired by imaging a patient with a phantom object in proximity to the patient.
12 . The method according to claim 11 , comprises, before detecting and segmenting and after providing one of the at least one radiographic image as a candidate image for classification, selecting the machine learning model to be applied based on said candidate image for classification, comprising:
determining that the machine learning model to be applied is a first machine learning model being generated with clinical images if the candidate image is a clinical image;
determining that the machine learning model to be applied is a second machine learning model being generated with phantom images and/or mixed images if the candidate image is an image of a phantom object or a mixed image of a patient with a pocket phantom.
13 . The method according to claim 10 , wherein said machine learning model has been generated by implementing a method for generating a machine learning model according to claim 1 .
14 . The method according to claim 10 , further comprising, before detecting and segmenting at least one ROIs as candidate reference ROI, pre-processing said candidate image by using at least one of the following image processing techniques: denoising, resampling, diffusion-weighted magnetic resonance imaging and apparent diffusion coefficient mapping.
15 . The method according to claim 10 , wherein detecting and segmenting in the candidate image at least one regions of interest as candidate reference ROI comprises, if it is determined that the machine learning model to be applied is the first machine learning model, detecting and segmenting in the candidate image at least one non-pathological regions of interest as candidate reference ROI.
16 . The method according to claim 15 , wherein detecting and segmenting in the candidate image at least one non-pathological regions of interest as candidate reference ROI comprises segmenting various non-pathological anatomical structure of interest, and segmenting and detecting ROIs in the segmented ASIs as candidate reference ROIs.
17 . The method according to claim 10 , wherein detecting and segmenting in the candidate image at least one region of interest as candidate reference ROI comprises, if it is determined that the machine learning model to be applied is the second machine learning model, segmenting various phantom-derived structure of interest, and segmenting and detecting ROIs in the segmented PDSIs.
18 . The method according to claim 10 , further comprises, after detecting and segmenting in the candidate image at least one region of interest as candidate reference ROI, optimizing at least one of the previously detected and segmented candidate reference ROIs by at least one image processing technique chosen among binning of image values and rescaling.
19 . The method according to claim 10 , wherein identifying valid candidate reference ROIs from a plurality of candidate reference ROIs further comprises comparing the measured radiomics values to a predetermined range of radiomics values chosen from a set of reference value ranges and determining a valid candidate reference ROI if the measured radiomics values thereof fall into the predetermined range of radiomics values.
20 . The method according to claim 10 , wherein comparing the measured radiomics values of each said valid candidate reference ROIs to each of the radiomics value of each of said at least two reference clusters comprises calculating distances between the measured radiomics values of each of the valid candidate reference ROIs and each of the radiomics values of each of the at least two reference clusters.
21 . The method according to claim 10 , wherein classifying the candidate image into one of the at least two reference clusters based on the comparison comprises selecting the cluster having the smallest calculated distance and/or eliminating the cluster having the largest calculated distance.
22 . A computer program product comprising instructions which, when implemented by at least one digital processing device, performs at least ones among the method according to claim 1 and the method according to claim 10 .
23 . A method for generating a machine learning model for automatic classification of radiographic images acquired by various acquisition protocols, comprising:
Providing a plurality of radiographic images of at least two different patients, the radiographic images of each of the at least two different patients being acquired with acquisition protocols different from each other,
Detecting and segmenting in each of said radiographic images at least one regions of interest as reference ROI,
Measuring at least one radiomic feature per reference ROI,
Identifying valid reference ROIs from a plurality of reference ROIs based on the measured radiomics values,
Clustering the measured radiomics values of said valid reference ROIs into at least two reference clusters according to a set of characteristics of image acquisition,
wherein detecting and segmenting in each of said radiographic images at least one regions of interest as reference ROI comprises, if the radiographic images are phantom images and/or mixed images each combining a patient and a pocket phantom, segmenting various phantom-derived structure of interest based on a reference material, and segmenting and detecting ROIs in the segmented PDSIs.
24 . A method for generating a machine learning model for automatic classification of radiographic images acquired by various acquisition protocols, comprising:
Providing a plurality of radiographic images of at least two different patients, the radiographic images of each of the at least two different patients being acquired with acquisition protocols different from each other,
Detecting and segmenting in each of said radiographic images at least one regions of interest as reference ROI,
Measuring at least one radiomic feature per reference ROI,
Identifying valid reference ROIs from a plurality of reference ROIs based on the measured radiomics values,
Clustering the measured radiomics values of said valid reference ROIs into at least two reference clusters according to a set of characteristics of image acquisition,
wherein identifying valid reference ROIs from a plurality of reference ROIs further comprises comparing the measured radiomics values to a predetermined range of radiomics values chosen from a set of reference value ranges and determining a valid reference ROI if the measured radiomics values thereof fall into the predetermined range of radiomics values.