| Sonstiges: |
- Nachgewiesen in: USPTO Patent Applications
- Sprachen: English
- Document Number: 20180276821
- Publication Date: September 27, 2018
- Appl. No: 15/989457
- Application Filed: May 25, 2018
- Claim: 1. A method for automatically recognizing liver tumor types in ultrasound images, comprising: S1: using a model to represent a case, and using a local classifier to represent possible change forms of a lesion; S2: inputting ultrasound images of a group of cases and the type of liver cancer in each case as a training sample; S3: initializing the values of model parameters all to 0, or randomly initializing the values with a Gaussian probability distribution which has the expectation of zero; S4: based on the model parameters, in an ultrasound image video of a case, using a dynamic programming algorithm to search for an optimal Region of Interest (ROI) location, size, and time for each local classifier, so that when a model determines that the lesion class of the training sample is correct, a maximum score is obtained; S5: using a graph cut algorithm to determine a specific change form and ROI of the case; S6: based on the ROI determined in S5, using the training sample in S2 as an input, using a cutting-plane algorithm to train it, and using outputs of the algorithm as model parameters to obtain possible locations and sizes of lesions in the case and the number of local classifiers; S7: repeating S4 to S6 to acquire the lesion type of each case of a training sample data type, judging the correctness of the acquired lesion type, and when the number of judgment errors is fixed or the number of repetition steps reaches a preset value, obtaining training model parameters; S8: using the training model parameters acquired in S7 to determine an optimal ROI location, size, and time for all local classifiers in an ultrasound image video of a case to be detected, using a graph cut algorithm to determine a specific change form and ROI of the case to be detected, and using a model to obtain a probability score for a lesion according to the determined change form and ROI; and S9: repeating S2 to S8 for all lesion types in the case to be detected, and obtaining a probability score for one lesion by each repetition, a lesion type corresponding to the highest probability score being the lesion type of the case to be detected.
- Claim: 2. The method according to claim 1, wherein an ROI is respectively extracted for an input case ultrasound image at three phases namely an arterial phase, a portal phase and a delayed phase, the three ROIs being used to express FLLs; appearance features of an ROI interior, appearance features of an ROI boundary and appearance features of an ROI periphery are respectively extracted on the ROI interior, the ROI boundary and the ROI periphery of the ROI at each phase, and an average grayscale difference between the ROI interior and the ROI boundary and an average grayscale difference between the ROI interior and the ROI periphery are also acquired in the ROI at each phase; and when regional features of the ROI interior, the ROI boundary and the ROI periphery are extracted, the contrast, correlation, energy, and identity of a regional grayscale co-occurrence matrix are used as appearance features.
- Claim: 3. The method according to claim 2, wherein in S3, pruning is used first, and then an optimal ROI is searched by using a dynamic programming method, a specific process being as follows: a pruning process comprises time and space pruning; the time pruning is: calculating the grayscale feature of a grayscale co-occurrence focus in a certain frame in an ultrasound image video, and making a difference between vector frames to obtain a change value used to represent each frame; using a series of values obtained within a period of time to represent the degree of change of each frame within the period of time; arranging this series of values in a time sequence, and extracting the local maximum points to obtain a frame most dramatically changing within a period of time; selecting these frames corresponding to the local maximum points to obtain candidate frames of a lesion region; and in the remaining candidate frames, removing some unimportant regions based on experience; the space pruning is mainly achieved by using a priori of significance and location, comprising: calculating a saliency map of the entire image first, normalizing the saliency map, and averaging salient values in a region to obtain a salient value of the region, the priori of location being image-centered Gaussian distributions; and multiplying the values obtained from priori information of these two parts to obtain the probability that a certain region is an FLL region, and selecting a region of the probability is greater than the threshold as a candidate ROI; dynamic programming is used to search an ROI; after time and space pruning, a different number of frames will be retained within three different phases, and a different number of candidate ROIs will be retained within each frame; in the timing constraint, two ROIs adjacent to each other in a chronological order must be close to each other at a spatial location; the dynamic programming method is: finding an ROI within a previous phase, so that the ROI is in the vicinity of a current ROI, and its score plus timing relationship score of the current ROI can reach a maximum value; and searching a local classifier, so that an appearance score of the current ROI is maximized
- Claim: 4. The method according to claim 3, wherein in S3, after searching for an optimal ROI, selection of the local classifier for the case will also be adjusted, the adjustment mode being: increasing, decreasing, or maintaining the number of local classifiers; in this process, the goal is to maximize a local classifier score selected for each case while maximizing the similarity measure of two ROIs under the same local classifier.
- Claim: 5. The method according to claim 4, wherein a specific mode to achieve the above goal is: using a Euclidean distance between two ROI features as a measure of similarity therebetween, and using a graph cut algorithm to solve the problem by converting into a graph label problem.
- Current International Class: 06; 06; 06; 06; 06; 06; 06; 61; 61
|
