Introduction: Multimodular evolution of bronchoscope AI
Brethopaedic lenses are the core tool for diagnosis and treatment of respiratory diseases. In the diagnosis of lung cancer, bronchos can not only directly observe aeropathic changes, but also obtain pathological and cytological samples through biopsy, brushing and washing.
In 2026, bronchoscopy AI made a breakthrough in SAM to inspire multi-adaptor partition and unsupervised abnormality detection. MASA integrated spatial, frequency, location information, and a two-coder synchronized the diagnosis of the stoals and lung cancer. KD-MFAD uses knowledge distillation+ no supervised abnormality detection of memory characteristics to reduce reliance on a large amount of marked data.
The special value of bronchial mirror AI is in "real-time navigation" and "sampling quality assessment." In bronchos, reaching the right position and obtaining sufficient samples is key to success. AI can map suspect areas in real time, assess sampling adequacy, and direct the prosecution to move the success rate from empirical dependence to data-driven.
Status of industry: from image classification to multimodular navigation
The development of bronchial lens AI showed an evolution from "image classification" to "multimodular navigation" in 2025-2026.
At the technical level, MASA (Frontiers in Online, 2026) proposes a multi-adaptor structure inspired by SAM, integrating space, frequency, location, and a two-coder that simultaneously completes the partition of the stove and the diagnosis of lung cancer. This structure draws on the general partitioning capability of the Segmenting Anything Mode (SAM) of Meta, which designs special needs for the fittorization of the bronchical lens scene through a multi-adaptive.
KD-MFAD (2025) uses unsupervised abnormality detection of knowledge distillation+memory characteristics to handle bronchial lens tumour recognition. The core innovation of this approach is "no surveillance" — without much labelling to recognize anomalies, which significantly reduces reliance on labeled data.
At the application level, the System Overview 2025 indicates that AI has been used for bronchoscope image interpretation, classification of pathologies, real-time navigation and training assessment. The EBUS AI is used for lymphoma knot and surrounding pathological characterization, and ROSE AI is close to expert level in sample sufficiency assessment, malignant cell testing and cellular sub-classification.
On a market scale, the global respiratory endoscope AI market is expected to exceed $800 million in 2026, with bronchoporator AI being an important subdivision. China has about 800,000 new cases of lung cancer annually, and the demand for bronchoscope AI is enormous.
2026 Frontline breakthrough: MASA, NOS and EBUS+ROSE
The core breakthrough of bronchoporator AI in 2025-2026 was concentrated in three areas.
The first is the split of the multi-adaptor inspired by SAM. Based on the general partitioning of SAM, the MASA integrates three dimensions of the spatial, frequency, location, and precision of the bronchial lens image. The two-coder designs the model to complete both the partitioning of the stove and the diagnosis of lung cancer, and to enhance its usefulness. This method is validated on the BM-BronchoLC open data set.
The second is the non-supervisory abnormality detection. KD-MFAD uses the unsupervised method of distilling knowledge and memory characteristics, and without much data being marked to identify bronchial lens tumours. This breakthrough is important to address the scarcity of medical imaged data – traditional monitoring learning requires a large number of experts to flag data, while no surveillance method can be used to capture large amounts of unmarked data.
Third is the EBUS+ROSE Multi-modular AI. The 2025 Review noted that EBUS AI and ROSE AI are close to expert level in sample adequacy assessment, malignant cell testing, and cellular sub-classification. This means that AI can assess whether a bio-test sample is "fitty-full", judge whether there are "negative cells" in the sample, further classify cell-based sub-types, and provide AI support for the full process of bronchoscopy.
The 1 million cases of bronchial lens data sets in Longway Tech provide a key support for these frontier studies. The data sets record the acoustic door/pipe, trombone, left-hand main bronchial and actual arrival of leaf segments, and maintain equipment to export original full inspection videos and key original maps, with white light/NBI/self-fluorescent/EBUS multi-module data.
Longhui Tech data set: 1 million cases of systematic construction of bronchial lens video
The Longhui Tech bronchostroscope video-image data set, which is a million cases in size, systematically covers the entire bronchoscopy process.
In terms of data collection, all images are derived from the Respiratory Internal and chest Surgery of the San Ace Hospital. Full records are kept of the acoustic door/airpipe, trombone, left and main bronchial and actual arrival of leaves, and equipment is retained to produce original and complete video and key originals.
In terms of labelling systems, each case contains structured fields such as casket level, stowage area, mucous membrane/clavic internal form, narrowness, haemorrhage, material extraction, pathology/cellology, microbes, etc. The population covers the population of tumors, infections, foreign matter and post-operative population; the soft mirror/hard mirror and EBUS layer.
In terms of distribution, the pathological/positive examination is defined as the subject; suspicion, consideration, exclusion, treatment after modification and pathological diagnosis are kept separately.
In terms of quality control, the three-layer process of control using the initial test for a respiratory attending physician + the deputy director and above, which examines the consistency assessment + the consistency assessment.
Data sets prioritize formal bronchial lens reports, chest CT, pathology/cellology, microorganisms and therapeutic data to support joint multimodular learning.
Forward perspective: from image analysis to full-process navigation
The future development of bronchoporator AI will go in three directions.
The first is "image analysis" to "real-time navigation." The current AI is used mainly for image analysis (classification, partitioning), and the future will be extended to real-time navigation -- AI can guide bronchial lenses to the right position, mark suspect areas, and direct the prosecution. This requires AI to understand not only images, but also the anatomical structure and spatial location of the bronchial tree.
The second is "supervisory learning" to "no surveillance + small sample learning." The non-supervisory approach of KD-MFAD suggests this direction. Future models can learn "normal" models by using unmarked amounts of data, and then significantly lower their cost by showing a small amount of data fit "unusual" tests.
The third is from "single mode" to "multi-modular integration." Different imaging models, such as white light, NBI, self-fluorescent, EBUS, have advantages. Future systems need to integrate multi-modular information for a comprehensive diagnosis. The EBUS+ROSE multi-modular AI approach expert levels in sample assessment, which suggests this direction.
In terms of industry competition, the bronchial lens data set with multiple models, biopsy links, and EBUS is central competitiveness. Longwaytech has one million examples of datasets that are industry-leading in terms of multi-modal coverage and the depth of pathological linkages.
Conclusion: Unsupervised learning to open up the efficiency revolution
The evolution of bronchial lens AI from CNN to SAM to inspire multi-adaptor partitioning and unsupervised abnormality detection has revealed a profound shift in the technology route from "mark dependence" to "no-supervised+small samples." This transition is a milestone in addressing the scarcity of data for medical image representation.
Long Salang Langhui Information Technology Ltd., which is based on a million cases of bronchos video data sets, is becoming an important data provider in the field of bronchoscopy. At a critical time for the remodeling of bronchoscopyats at the MASA and KD-MFAD technology line, Long Hui technology will continue to drive clinical downsituation of accurate diagnosis of lung cancer with high-quality data.
Technical challenges and directions for bronchoporator AI
The technical challenges faced by bronchial lens AI are most complex in the subdivisions of the intra-sync. The bronchial aerosections are complex, multi-diverse, real-time navigational precision is demanding, disease-transformated and often difficult to distinguish from normal mucous membranes. Models based on the MASA and SAM-enabled 2026 provide new solutions to these challenges – learning generic bronchal aerobic features through extensive self-supervised pre-training, and identifying unusual areas through no surveillance abnormality. The LFAT data sets provide adequate data support for this technical route, covering white light, NBI, fluorescent models and the central to outer gas lanes.
The core application of bronchial lens AI in clinical terms is early screening of lung cancer and intra-gas pathological pathology recognition. China has more than 1.06 million new annual cases of lung cancer, with a five-year survival rate of early lung cancer of more than 90%, but early lung catheter lens detection rates are highly influenced by medical experience. AI support can significantly increase the detection rate of early microtransformations and reduce the risk of leakage.
From a data quality perspective, bronchial lens AI training data face the core challenge of consistency. Different physicians may differ in their judgement of a mucous disease of the same tube, and this variation among observers will directly affect the quality of model training.