Search

CN-122023433-A - Medical image segmentation method based on consistent quality control and closed-loop constraint

CN122023433ACN 122023433 ACN122023433 ACN 122023433ACN-122023433-A

Abstract

The invention discloses a medical image segmentation method based on consistent quality control and closed-loop constraint, which comprises the steps of obtaining medical image data to be segmented and inputting a segmentation network to obtain an initial segmentation mask, applying topological and morphological closed-loop constraint to the initial segmentation mask to output a morphological rationality index set, applying various equivalent transformations to the medical image data to form a sample set, reasoning the sample set to obtain a mask set, calculating to obtain consistent quality control scores based on the mask set, triggering an automatic error correction reasonement strategy library to obtain a final segmentation mask when the consistent quality control scores are lower than a preset threshold or the morphological rationality index set does not meet preset constraint conditions, aligning the inter-period image data of the same patient and calculating the variation of the final segmentation mask, and applying follow-up consistency constraint to generate a review prompt or follow-up variation report and output.

Inventors

  • XIE SHIPENG
  • CHEN XIAOYU

Assignees

  • 南京邮电大学

Dates

Publication Date
20260512
Application Date
20260129

Claims (9)

  1. 1. The medical image segmentation method based on consistent quality control and closed-loop constraint is characterized by comprising the following steps of: acquiring medical image data to be segmented; Inputting the medical image data into a segmentation network to obtain a probability map of a lesion or structure and an initial segmentation mask; applying topology and morphology closed-loop constraints to the initial segmentation mask to output a morphology rationality index set; applying a plurality of equivalent transformations to the medical image data to form a sample set, and reasoning the sample set to obtain a mask set; based on the mask set, calculating result consistency and obtaining consistency quality control scores; When the consistency quality control score is lower than a preset threshold value or the form rationality index set does not meet a preset constraint condition, triggering an automatic error correction re-reasoning strategy library to update reasoning parameters or local re-reasoning so as to obtain a final segmentation mask; Aligning the medical image data of the baseline period and the image data of the follow-up period of the same patient, calculating the variation of the final segmentation mask, and applying follow-up consistency constraint based on the consistency quality control score to generate a review prompt or a follow-up change report; outputting the final segmentation mask, the consistent quality control score, the review prompt or the follow-up change report.
  2. 2. The medical image segmentation method based on consistent quality control and closed loop constraints of claim 1, wherein the medical image data is two-dimensional slice or three-dimensional volume data.
  3. 3. The medical image segmentation method based on consistent quality control and closed loop constraints of claim 1, wherein the medical image data is resampled to uniform spatial resolution and intensity of the medical image data is normalized or window-level mapped prior to input into a segmentation network.
  4. 4. The method of medical image segmentation based on consistent quality control and closed loop constraints of claim 1, wherein the step of applying topological and morphological closed loop constraints to the initial segmentation mask to output a set of morphological rationality indicators comprises: calculating morphology indexes of the initial segmentation mask, wherein the morphology indexes comprise the number of connected domains, the maximum connected domain occupation ratio, the number of holes, the boundary length, the compactness, the slender ratio and the distance between the boundary and an organ; and comparing the morphological indexes with a preset index threshold value to obtain a morphological rationality index set.
  5. 5. The medical image segmentation method based on consistent quality control and closed loop constraints of claim 3, wherein the set of morphological rationality indicators comprises a number of connected domains and a maximum connected domain duty cycle, a number of holes and a hole area duty cycle, a compactness, an elongation ratio, a boundary smoothness index, and an organ boundary distance.
  6. 6. The medical image segmentation method based on consistent quality control and closed loop constraints of claim 1, further comprising collision detection, the collision detection comprising the steps of: Analyzing image report text or structured examination information corresponding to the medical image data to obtain a weak supervision constraint set; And aligning the weak supervision constraint set with a measurable index of the initial segmentation mask, and marking the initial segmentation mask as reporting conflict when the measurable index conflicts with the weak supervision constraint set.
  7. 7. The medical image segmentation method based on consistent quality control and closed loop constraints of claim 5, wherein the collision detection further comprises using the labeled initial segmentation mask as a training or online correction for loss terms of the segmentation network.
  8. 8. The method of medical image segmentation based on consistent quality control and closed loop constraints of claim 1, wherein the step of calculating result consistency and obtaining consistent quality control scores based on the set of masks comprises: Calculating to obtain the average overlapping degree and the average boundary distance of the mask set; and normalizing and fusing the average region overlapping degree and the average boundary distance to obtain the consistent quality control score.
  9. 9. The medical image segmentation method based on consistent quality control and closed loop constraints of claim 1, wherein the step of aligning the medical image data of the baseline period and the image data of the follow-up period of the same patient and calculating the variation of the final segmentation mask while applying the follow-up consistency constraint based on the consistent quality control score to generate a review prompt or a follow-up variation report comprises: the trans-phase medical image data of the same patient is aligned based on the coarse alignment of the organ ROI and the fine alignment based on mutual information or feature matching; Calculating a final segmentation mask of the inter-period medical image data of the same patient to obtain the variation; obtaining a variation reliability based on the alignment quality index, the layer thickness difference, the noise level and the consistency quality control score calculation; When the change reliability is low or the change conflicts with the morphological rationality index set, judging that the pseudo change is at high risk and generating the rechecking prompt area; the follow-up change report is generated when the change confidence is high and the change is concentrated in a high confidence region.

Description

Medical image segmentation method based on consistent quality control and closed-loop constraint Technical Field The invention relates to the technical field of medical image segmentation, in particular to a medical image segmentation method based on consistent quality control and closed-loop constraint. Background Along with the continuous improvement of the resolution of medical imaging equipment and the promotion of informationized construction of hospitals, the image data such as CT, MRI, ultrasound and the like have more remarkable roles in clinical screening, diagnosis and staging, preoperative planning, drawing of a radiotherapy target area and follow-up evaluation. In order to rapidly extract organ structures and lesion areas from massive image data, medical image segmentation techniques are widely studied and gradually applied to clinical auxiliary scenes. In the existing method, manual or semi-automatic sketching is still a common means of many departments, but the mode has strong dependence on doctors' experiences, and sketching layer by layer on a three-dimensional image is time-consuming and labor-consuming, and is easily influenced by differences of workload, film reading habit and boundary understanding, so that the consistency and repeatability of segmentation results are insufficient. In recent years, a deep learning segmentation model has a better effect on a public data set, so that automatic segmentation is possible, but the methods still face the problem of insufficient stability under the complex condition of real clinical data, especially in the phenomena of fuzzy boundary, changeable focus form, low contrast scene, partial missing, over-segmentation or contour dithering and the like easily occur in model output, and further popularization and application are limited. From the data and supervision perspective, high-quality labeling of medical image segmentation often requires a doctor with professional qualification to carry out fine pixel-by-pixel and voxel-by-voxel delineation, and has high cost and long period, and unavoidable subjective differences exist in judgment of focus boundaries by different doctors, so that supervision signals have noise and uncertainty. Meanwhile, the data form actually accumulated clinically presents the characteristics of more images and less fine marks, and in more cases, descriptive information or measuring information in an image report exists only and a pixel-level label which can be directly used for training is absent. The existing multiple segmentation methods still mainly rely on strong supervision training, and it is difficult to fully utilize the weak tag information which is easier to obtain. Under weakly or unsupervised conditions, the model is prone to degradation to bias predictions that dominate the background, or to false dependencies on certain high frequency textures/artifacts, thus reducing generalization ability. Because different hospital equipment, scanning protocols, reconstruction kernels and layer thickness parameters have large differences, performance fluctuation of the model also often occurs when the model is applied across domains, and the risk of clinical deployment is further aggravated. In addition, the existing medical image segmentation system generally lacks quality control and closed-loop engineering mechanisms, the common flow is mostly that mask results are directly output after one-time reasoning, quantitative evaluation on the reliability of the results cannot be given, and an automatic error correction and re-reasoning strategy when the results are abnormal is also lacking. For clinic, whether the segmentation result is reliable or not is often more critical than the average precision index, namely, when the model outputs an error mask on noise, motion artifact or low-contrast images, if the system cannot prompt risk or trigger error correction in time, the deviation of subsequent measurement, stage evaluation and treatment decision can be caused. Especially in follow-up scenes, the images of the same patient in multiple phases are affected by factors such as respiratory phase, body position change, layer thickness difference and the like, and pseudo-change is easy to occur by simply comparing two-phase masks, so that judgment of tumor progress or curative effect is interfered. Therefore, a medical image segmentation scheme is needed, which can provide consistent quality control scoring in a segmentation reasoning stage, form a data closed loop by using report weak supervision information in a training stage, introduce topological rationality constraint in an output stage and have automatic error correction and re-reasoning capability, and simultaneously has span consistency and pseudo-change inhibition mechanism for a follow-up scene, so that the reliability, the interpretability and the floor-standing performance of the system in a real clinical environment are improved. Disclosure of Invention In order to ove