Breast cancer is the leading cause of death among women worldwide, and China is one of the countries with the fastest-growing incidence rate, increasing by 2% annually. After years of persistent efforts to develop and utilize advanced technologies and drug treatments, it was discovered that the prognosis for most late-stage cancer patients remains largely unknown. Humanity has finally had to accept the option of moving the treatment point forward, focusing on early diagnosis, early detection, early prevention, and early treatment. Therefore, in response to the threat of breast cancer, the country launched a national free public health program on screening for two cancers (cervical and breast cancer) in 2009.`nbsp_tag
Physiological Differences Prevent China from Simply Copying the Western Approach
Although the country has launched a free public service program for screening for two cancers, the progress has been unsatisfactory due to factors such as a lack of public awareness of medical health, immature technology, and high labor costs. Currently, the most reliable method for early screening of breast cancer is medical imaging. Developed countries mostly use mammography as a routine examination method. American family doctors recommend mammography screening every 2 years for women aged 40-49, and recommend screening based on individual circumstances for women over 50, while recommending mammography screening every 2 years for women aged 50-74. China also clarified in the "Guidelines for Breast Cancer Screening, Early Diagnosis, and Treatment" issued in 2018 that mammography combined with ultrasound should be used for breast cancer screening.
Due to the physiological structural differences between Asian women and Caucasian women,the difficulty of mammographic imaging in identifying early breast cancer in Asians is far greater than that in Caucasiansandultrasound screening for breast cancer has low sensitivity and poor specificity and only serves as an auxiliary examination method for mammography. However, in the Asia-Pacific region, ultrasound as the primary screening method may be more suitable. However, ultrasound is a dynamic image examination, and it places high demands on the skill and focus of the examiner, so using ultrasound as the main breast cancer screening method also presents difficulties.
Ultrasound Breakthrough`nbsp_tagTechnical Challenges
Since ultrasound screening has drawbacks, can we use some schemes to supplement it? Obviously, AI systems can better free up doctors, reduce their workload, and, based on big data, make the examination results more accurate. However, the traditional AI analysis method of intercepting images and then using AI for analysis is not very meaningful for ultrasound examinations.
The development of AI technology in the field of medical image-assisted diagnosis, along withthe maturity of algorithms and dataand other factors, continues to evolve.Initially, AI primarily targeted 2D imagingbecause the openness of algorithms and data was divided into two stages. The first stage involved products with publicly available datasets such as lung nodules/fundus, and the second stage involved disease types with no publicly available datasets, such as breast mammography. Since then, AI-assisted diagnostic products for 3D imaging have been developed and widely used in fields such as CT/MR that require three-dimensional reconstruction, although this greatly increases the difficulty of AI analysis, the effect is excellent in current applications.
However, AI recognition is still limited to the analysis of static images. For ultrasound, which involves a series of continuous video signals, no AI product has been able to achieve this before because in this field,there are no publicly available datasets or ready-made algorithm modelsand there is not even a recognized method for data acquisition and annotation. Because two-dimensional images are relatively easy to obtain and the processing algorithms are less difficult, the exploration of ultrasound problems at the national and global levels has mostly focused on two-dimensional imaging. However, for practical application scenarios, the information on lesions in two-dimensional images is insufficient, and the accuracy of the analysis has an inherent bottleneck, which is a significant gap from true clinical auxiliary diagnosis. This means thatultrasound AI-assisted diagnostic products not only need to design new methods for the acquisition, storage, and recognition of video data but also need to design a specialized algorithm。
Timeliness and Accuracy Conflicts
The main technical challenge currently facing ultrasound R&D is the high requirements for both real-time performance and accuracy, that is, both real-time performance and extremely high accuracy are required. CT/MR image acquisition and diagnosis are performed separately, so the AI system has sufficient time to process the images. However, due to the diagnostic process and characteristics of ultrasound, it requires ultrasound AIto achieve real-time auxiliary diagnosis. AI needs to view the scanned video images in real-time while the doctor is operating the ultrasound instrument, give a diagnosis, and then feedback the results to the doctor in order to effectively assist the doctor in making a diagnosis.
Real-time analysis places higher demands on the algorithm and computing power of AI, especially in the medical field, where the requirements for accuracy are exceptionally high, which means that the complexity of the model will be greatly increased to extract richer features. However, ultrasound AI cannot sacrifice time to gain accuracy, so there are huge challenges in algorithm design. At the same time, due to the huge amount of video data, daily ultrasound examinations do not store a large amount of video data; usually, only a few representative images are kept, which poses a great challenge to ultrasound data acquisition.
In addition, excellent models also rely on high-quality labeled data. For ultrasound video data, each patient usually takes about ten minutes, with30 frames per secondcalculated, then the two-dimensional ultrasound images of one patient are10*60*30=18000These images need to be annotated by experts with extensive knowledge of ultrasound medicine. The current situation in China is that there is already a shortage of nearly 100,000 ultrasound doctors, and their daily work is very busy. Finding professional personnel to complete such a monumental annotation task is very challenging.
Creating New Models: Medical Precision Intelligent Breast Ultrasound Intelligent Detection System
As a specific part of ultrasound detection, the characteristics of the breast itself also increase the difficulty of AI technology R&D. This is mainly manifested in two aspects.First, it is difficult to distinguish between breast lesions and fat; second, it is difficult to judge the benign and malignant nature of breast lesions.
Many fat sections are similar to lesions, so multiple-angle scanning is needed during ultrasound examination to reduce false-positive diagnoses. For video detection algorithms, due to the need for real-time performance, it mainly focuses on the currently received images and analyzes them. However, common algorithms can only recognize and analyze the lesion part and cannot recognize the surrounding information, unable to comprehensively judge from multiple angles like a real doctor. Therefore, the R&D path of breast ultrasound video detection and classification algorithms is quite tortuous.
Yizhun Intelligent's R&D team overcame the above-mentioned difficulties and releasedChina's first intelligent breast ultrasound detection systemYizhun Intelligent established a new model for this system, making targeted adjustments to the two technical difficulties of breast ultrasound. After the release of the "Guidelines for Breast Cancer Screening, Early Diagnosis, and Treatment in China," Yizhun Intelligent started to develop AI for breast cancer screening. In 2019, it pioneered the development of the Yizhun Intelligent Breast Mammography Intelligent Detection System. In 2020, Yizhun Intelligent made a breakthrough from AI recognition of 2D images to dynamic images, conquering the heights of ultrasound AI.
-
Addressing the problem of increased false positives due to the difficulty of distinguishing breast lesions from fat: Based on the multi-scale features of FPN, time-dimension information was added, and 3D CNN was used to extract three-dimensional features of the lesion in both time and space dimensions. By combining these features, false positives such as fat were effectively eliminated.
-
Addressing the problem of difficulty in judging the benign or malignant nature of lesions: To obtain more accurate benign-malignant classification results, first, the three-dimensional structure of the lesion was restored from videos of various cross-sections of the lesion; second, an attention mechanism was added to incorporate the features of the surrounding tissues into the lesion features with different weights, making full use of the lesion and its surrounding information to achieve more accurate classification results.
Yizhun Intelligent Breast Ultrasound Intelligent Detection SystemNo equipment modification is required, and no adjustment to the existing workflow is neededWhile the doctor is scanning the patient,the AI server performs real-time analysis and provides marking prompts on the interfaceIt can accurately capture lesions that flash for only milliseconds, effectively avoiding misdiagnosis due to physician visual fatigue and insufficient visual sensitivity.
As a clinically applied system, the Yizhun Intelligent Breast Ultrasound Intelligent Detection System also has the following five highlights:
-
Fast computing speed, low latency;
The system uses a neural network architecture search method (NAS) and RTX2080Ti. The processing speed reaches > 50 frames per second, and the detection result delay is <0.09 seconds, allowing for accurate capture of lesions that flash for only milliseconds.
-
High lesion identification and detection rate;
By simulating the restoration of 3D through convolutional neural network feature fusion, the benign or malignant nature of the lesion is determined.
-
Low false positives (false alarms);
The system effectively reduces the false positive rate by screening meaningful frames in all images for judgment.
-
Intelligent segmentation, automatic measurement;
Compared to using only the features of a certain cross-section of the lesion for attribute analysis, this system can also analyze the entire video containing all the information of the lesion, making full use of the information from various cross-sections of the lesion for a more comprehensive attribute analysis of the lesion as a whole, while also providing quantitative information such as the largest cross-sectional area, length, and width.
-
Structured report generation
One-click access, accelerating imaging report diagnosis efficiency; intelligent ultrasound findings and diagnosis, assisting clinicians in obtaining detailed text data of the disease.
The breast ultrasound intelligent detection system developed by Yizhun Intelligent may fundamentally change breast cancer screening and even the landscape of the entire AI-assisted diagnosis industry. Previous AI products mostly relied on radiology departments and used imaging equipment such as DR, CT, and MR. However, most radiology equipment cannot move the examination environment forward, while ultrasound equipment has unique advantages in this aspect. The small, radiation-free ultrasound can break free from the constraints of the department and move the examination environment forward, making breast cancer screening more convenient.
In addition, all previous AI-assisted diagnostic products used static image capture for AI analysis, whileAI directly analyzes dynamic imagesYizhun Intelligent's breast ultrasound intelligent detection systemcompletely subverts the inherent model of the AI-assisted diagnosis fieldunlocking new skills for medical AI.
According to feedback from early cooperating hospitals, ultrasound physicians can use the Yizhun Intelligent Breast Ultrasound Intelligent Detection System without changing their operating methods. The AI system can immediately provide reference opinions and automatic annotations, providing strong support for the judgment of ultrasound physicians. This not only shortens the physician's judgment time but also reduces the physician's energy consumption and eliminates the need for dedicated data entry personnel. At the same time, it significantly improves the accuracy of breast screening, significantly reduces the workload of ultrasound physicians, and saves manpower costs for hospitals, making the product widely loved by ultrasound doctors.
Finally, it is understood that Yizhun Intelligent is exploring the possibility of integrating AI systems and imaging equipment into modular units. In the future, disruptive AI systems combined with modular imaging products will inevitably cover more disease examinations. China's two-cancer screening has been conducted for 10 years, but the number of breast cancer deaths has continued to rise during this period, largely because of late detection. The breast ultrasound intelligent detection system can help detect breast cancer earlier during screening, reducing breast cancer mortality.
Analysis, ultrasound, AI, lesion, intelligent, system, breast, diagnosis, breast cancer, imaging
