Lightweight nucleus segmentation system and method based on effective knowledge acquisition

Abstract

The invention discloses a lightweight nucleus segmentation system and method based on effective knowledge acquisition. The training of a teacher model is supervised by using a label of a cell nucleus segmentation data set; the trained teacher model teaches the image information capturing ability to the student model through a knowledge distillation method for obtaining effective knowledge; in the process, the label of the data set also participates in supervision when the student model is trained; the obtained student model can accurately segment and classify the cell nucleus on the pathological image. During cell nucleus segmentation, pathological images to be processed are placed in the same folder and sent into the model in the same batch, and then a corresponding cell nucleus segmentation result is obtained. The method can be operated on equipment with low computing power, and under the same hardware condition, the operation time is greatly shortened compared with other cell nucleus segmentation models, and meanwhile, excellent segmentation and classification effects are kept. Therefore, the method has practical application value and popularization prospect.

Description

Technical Field

[0001] The present invention relates to the field of medical image processing, and in particular to a lightweight cell nucleus segmentation system and method based on effective knowledge acquisition. Background Art

[0002] Segmenting the cell nuclei in pathological images and obtaining relevant morphological information can assist pathologists in diagnosing the status of cancer and the prognosis of patients.

[0003] The resolution of pathological images after scanning often reaches the million level, and the number of cell nuclei in an image is large, with different shapes and dense positions, so it is difficult to segment them manually.

[0004] With the development of deep learning technology, models for automatic segmentation and classification of cell nuclei in pathological images have also emerged.

[0005] However, in order to achieve excellent segmentation performance, these models are often designed to be bloated, which leads to high requirements for computer equipment when running these models. In addition, the model takes a long time to process pathological images. For example, the current mainstream cell nucleus segmentation model takes more than an hour to process a whole pathological slice, and it takes even longer on devices with less computing power.

[0006] This situation limits its widespread application in auxiliary diagnosis. Summary of the invention

[0007] The purpose of the present invention is to overcome the above-mentioned shortcomings and deficiencies of the prior art and to provide a lightweight cell nucleus segmentation system and method based on effective knowledge acquisition.

[0008] The present invention can also run on devices with lower computing power, and under the same hardware conditions, the running time is greatly reduced compared with other cell nucleus segmentation models, while maintaining excellent segmentation and classification effects. Therefore, the present invention has practical application value and promotion prospects.

[0009] The present invention is achieved through the following technical solutions:

[0010] A lightweight cell nucleus segmentation method based on effective knowledge acquisition includes the following steps:

[0011] S1. First, use the labels of the cell nucleus segmentation dataset to supervise the training of the teacher model;

[0012] S2. The trained teacher model is used to obtain effective knowledge through knowledge distillation, and the image information capture capability is taught to the student model. In this process, the label of the dataset is also used to supervise the training of the student model.

[0013] S3. The obtained student model is capable of accurately segmenting and classifying cell nuclei in pathological images.

[0014] When the cell nucleus is segmented, the pathological images to be processed are placed in the same folder and sent to the model in batches to obtain the corresponding cell nucleus segmentation results.

[0015] A lightweight cell nucleus segmentation system based on effective knowledge acquisition, including

[0016] Extract image information device;

[0017] Predicted nuclear background decoder;

[0018] Nuclear demarcation edge decoder;

[0019] Nucleus category decoder;

[0020] The results finally obtained by the predicted cell nucleus background decoder, the cell nucleus boundary edge decoder and the cell nucleus category decoder are applied to the watershed algorithm to generate a cell nucleus instance mask;

[0021] The predicted cell nucleus background decoder comprises two Dense Block decoder modules C and an image information decoding module A in sequence;

[0022] The cell nucleus boundary edge decoder sequentially comprises two decoder modules E containing Dense Block and curvature and an image information decoding module A;

[0023] The cell nucleus category decoder comprises two decoder modules D containing Dense Block and hole convolution and a decoder module B containing hole convolution in sequence.

[0024] Compared with the prior art, the present invention has the following advantages and effects:

[0025] The existing nuclear segmentation model has a long running time and high requirements on the computing power of the equipment, which limits the wide application and promotion of computer-aided diagnosis. The present invention provides an efficient lightweight cell nucleus segmentation model, using a lightweight network as a decoder to implement a decoder in a multi-task architecture.

[0026] The present invention adopts a knowledge distillation method to improve the accuracy of the model in segmenting and classifying cell nuclei by learning effective knowledge. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 The overall training architecture of the present invention, wherein the "student" refers to a lightweight kernel segmentation model.

[0028] Figure 2 Knowledge distillation for efficient knowledge acquisition.

[0029] Figure 3 The knowledge acquisition method corresponding to knowledge distillation.

[0030] Figure 4 The module structure used by the decoder in the cell nucleus segmentation system DETAILED DESCRIPTION

[0031] The present invention is further described in detail below in conjunction with specific embodiments.

[0032] The present invention discloses a lightweight cell nucleus segmentation method based on effective knowledge acquisition, comprising the following steps:

[0033] S1. First, use the labels of the cell nucleus segmentation dataset to supervise the training of the teacher model;

[0034] S2. The trained teacher model is used to obtain effective knowledge through knowledge distillation, and the image information capture capability is taught to the student model. In this process, the label of the dataset is also used to supervise the training of the student model.

[0035] S3. The obtained student model is capable of accurately segmenting and classifying cell nuclei in pathological images.

[0036] When the cell nucleus is segmented, the pathological images to be processed are placed in the same folder and sent to the model in batches to obtain the corresponding cell nucleus segmentation results.

[0037] The present invention discloses a lightweight cell nucleus segmentation system based on effective knowledge acquisition, including an image information extractor;

[0038] Predicted nuclear background decoder;

[0039] Nuclear demarcation edge decoder;

[0040] Nucleus category decoder;

[0041] The results finally obtained by the predicted cell nucleus background decoder, the cell nucleus boundary edge decoder and the cell nucleus category decoder are applied to the watershed algorithm to generate a cell nucleus instance mask;

[0042] The predicted cell nucleus background decoder comprises two Dense Block decoder modules C and an image information decoding module A in sequence;

[0043] The cell nucleus boundary edge decoder sequentially comprises two decoder modules E containing Dense Block and curvature and an image information decoding module A;

[0044] The cell nucleus category decoder comprises two decoder modules D containing Dense Block and hole convolution and a decoder module B containing hole convolution in sequence.

[0045] The present invention adopts a knowledge distillation method to obtain effective knowledge. That is, a model with a similar architecture is used as a "teacher". The model has a larger number of parameters and a more complex structure, so it can capture more image information. Through knowledge distillation, the powerful ability of the "teacher" to capture information can be imparted to the "student" (that is, a relatively lighter model will lack the ability to capture information). Use the "teacher" to guide the lightweight model ("student") we proposed. During the guidance, only the correct part predicted by the "teacher" is used to guide the "student", avoiding the teacher's own implicit wrong knowledge from misleading the "student's" learning. This helps students learn more correct knowledge.

[0046] As described above, the present invention can be better implemented.

[0047] The implementation methods of the present invention are not limited to the above-mentioned embodiments, and any other changes, modifications, substitutions, combinations, and simplifications made without departing from the spirit and principles of the present invention shall be equivalent replacement methods and shall be included in the protection scope of the present invention.

Claims

1. A lightweight cell nucleus segmentation method based on effective knowledge acquisition, Features The steps include: S1. First, use the labels of the cell nucleus segmentation dataset to supervise the training of the teacher model; S2. The trained teacher model is used to obtain effective knowledge through knowledge distillation, and the image information capture capability is taught to the student model. In this process, the label of the dataset is also used to supervise the training of the student model. S3. The obtained student model is capable of accurately segmenting and classifying cell nuclei in pathological images.

2. The lightweight cell nucleus segmentation method based on effective knowledge acquisition according to claim 1, Features When segmenting the cell nucleus, the pathological images to be processed are placed in the same folder and fed into the model in batches to obtain the corresponding cell nucleus segmentation results.

3. A lightweight cell nucleus segmentation system based on effective knowledge acquisition, Features include Extract image information device; Predicted nuclear background decoder; Nuclear demarcation edge decoder; Nucleus category decoder; The results finally obtained by the predicted cell nucleus background decoder, the cell nucleus boundary edge decoder and the cell nucleus category decoder are applied to the watershed algorithm to generate a cell nucleus instance mask; The predicted cell nucleus background decoder comprises two Dense Block decoder modules C and an image information decoding module A in sequence; The cell nucleus boundary edge decoder sequentially comprises two decoder modules E containing Dense Block and curvature and an image information decoding module A; The cell nucleus category decoder comprises two decoder modules D containing Dense Block and hole convolution and a decoder module B containing hole convolution in sequence.

Images