An acid-fast bacillus recognition method fusing dual prior constraints

CN122928843APending Publication Date: 2026-10-16TAIZHOU JIEWEI TIMES BIOTECHNOLOGY CO LTD
0 Cites 0 Cited by

Patent Information

Application Number
CN202611234199.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-14
Publication Date
2026-10-16

AI Technical Summary

Technical Problem

传统固定形态学后处理仅能推理过滤，无法反向参与模型训练，先验约束力度有限

Benefits of technology

[0050] This invention employs a network training method with a differentiable morphological prior operator to constrain the skeleton continuity and aspect ratio, thereby suppressing false positives from impurities such as mucus filaments and cell debris, and reducing false positives. By using a prior with a test interpretation standard as an independent loss term in the training, it supervises the grading and counting rules, ensuring the model's count distribution aligns with clinical grading rules. The output mean bacterial density directly matches laboratory reporting standards, reducing manual review and correction workload. A multi-branch joint supervision mechanism with two types of priors provides dual constraints on morphology and grading, addressing target morphology confusion through morphological constraints and counting bias through grading constraints, overcoming the limitations of single-prior schemes. Layered field-of-view sampling preprocessing during training automatically filters blank fields, ensuring the training sample distribution aligns with real clinical slide reading procedures, improving model generalization performance. A simple multi-branch network architecture and standardized output interfaces enable clinical system deployment, allowing direct adaptation to fully automated microscopic slide reading equipment in laboratories and tuberculosis clinics, with output recognition results meeting standardized clinical reporting requirements.

✦ Generated by Eureka AI based on patent content.
Patent Text Reader

Abstract

The present application relates to the technical field of medical microscopic image intelligent detection, in particular to an acid-fast bacillus recognition method fusing double prior constraints, comprising the following steps: collecting ZN-stained sputum smear oil immersion digital microscopic images, performing multi-focus layer fusion, Lab color channel dyeing correction and layered field screening pretreatment on the images, removing blank field images and retaining sputum-rich field to construct a sample pool; dividing the sample pool into a training set and a validation set, completing pixel-level labeling of the images and synchronously inputting clinical grading true value labels. The method synchronously embeds acid-fast bacillus morphological prior and clinical smear test interpretation standard prior into the network training process, constructs a multi-branch composite loss function to complete joint supervision, makes the network learning process simultaneously fit the inherent morphological characteristics of bacilli and the grading counting rules of the clinical laboratory, reduces the false detection probability of impurities, improves the detection capability of trace bacteria, and outputs results that adapt to the needs of clinical standardized reports.
Need to check novelty before this filing date? Find Prior Art