Yitu Healthcare's AI for Mammography Gains Clinical Traction: 16-Megapixel, Sub-Second Analysis Matches Senior Radiologists

Breasts have sparked countless wars and inspired endless praise, yet they are also particularly susceptible to tumors, such as breast cancer.

According to statistics from the National Cancer Center of China, breast cancer ranks first among all malignant tumors that women may develop, accounting for 16.51%. Globally, breast cancer is the most commonly diagnosed malignancy in women in both developed and developing countries.

For the early detection of breast cancer, clinical practice predominantly employs mammography. This modality is well-established, convenient, cost-effective, and demonstrates good reliability. Multiple medical guidelines assert that high-quality mammographic screening and follow-up can detect the majority of preclinical breast cancers, effectively reduce breast cancer mortality, and minimize unnecessary disability or avoid invasive treatments.

However, in China, determining who interprets the images has become the primary challenge. The interpretation of mammography is highly complex and requires a lengthy training period, resulting in an extreme scarcity of physicians specializing in this field, with only a few hundred nationwide. Meanwhile, the significant disparity in interpretive skills among physicians at different levels makes it difficult to ensure consistency in readings.

To address these challenges, Yitu Healthcare, a domestic AI company, has recently offered an alternative solution: AI-based image interpretation.

It is reported that the Intelligent Mammography Diagnosis System was developed and launched by Yitu Healthcare after several years of research and development. Leveraging robust algorithmic innovations and a vast dataset of tens of thousands of real-world mammographic images from multiple Grade 3A hospitals, this AI system achieves second-level interpretation of mammograms. It offers multiple functionalities, including glandular pattern classification, lesion detection, feature description, and intelligent BI-RADS categorization, and can automatically generate structured reports for use by radiologists.

In clinical practice across multiple Grade 3A hospitals, the system has demonstrated robust capabilities in lesion detection and learning. It not only detects breast masses, calcifications, architectural distortions, and asymmetries but also performs risk stratification for malignancy based on comprehensive lesion detection, thereby assisting physicians in identifying high-risk lesions.

The breast is beautiful, but it is not easy to clearly visualize internal lesions.

Unlike lung CT scans, which generate hundreds of images through layer-by-layer scanning to reconstruct the three-dimensional structure of the lungs, the naturally rounded physiological anatomy of breast tissue and the principle of vertical X-ray irradiation mean that routine bilateral mammography involves only two standard positions—mediolateral oblique (MLO) and craniocaudal (CC)—yielding a total of four single-image views. Consequently, compression paddles must be used to flatten the female breasts as much as possible, ensuring adequate separation of internal breast tissues, thereby producing mammographic images with optimal clarity to facilitate the identification of lesion locations.

However, this imaging modality must overcome the challenges of glandular obscuration and structural noise. It is akin to a hunter navigating a forest floor dappled with light, attempting to pinpoint prey in the treetops solely based on the patterns of light and shadow on the ground. Consequently, reconstructing 3D breast tissue from 2D imaging data places exceptionally high demands on the professional expertise of interpreting radiologists.

Compared with the more obvious calcifications and masses on imaging, architectural distortions and asymmetries, which account for approximately 30% of all lesions, are difficult to detect. Missed diagnoses are particularly likely when physicians lack sufficient experience or are fatigued.

Professor Peng Weijun, Head of the Breast Imaging Group of the Radiology Branch of the Chinese Medical Association and Director of the Department of Diagnostic Radiology at Fudan University Shanghai Cancer Center, has over 30 years of experience in interpreting mammograms. He stated that to reach a “senior” level, radiologists must possess a solid foundation in both anatomy and diagnostic imaging, strong spatial imagination, ample clinical experience, guidance from several outstanding mentors, and a growth period of 5–10 years.

The high standards for talent have made senior mammography interpretation experts extremely scarce. Across the Breast Imaging Group of the Chinese Medical Doctor Association’s Imaging Branch, there are only slightly more than 100 physicians considered senior specialists in mammographic interpretation, and among them, merely over 50 are recognized as “expert-level.” This scarcity forces these experts to shoulder the burden of interpreting mammograms for hundreds of millions of women in China who represent the potential demand for breast X-ray screening.

“Imaging features of breast lesions are less typical than those of pulmonary nodules, leading to substantial inter-observer variability in the diagnosis of many lesions. For instance, regarding architectural distortion, some experts may identify it as such, while others may not. Furthermore, the prevalence of dense breast tissue among Asian women further increases the likelihood of inconsistent interpretations.” Professor Peng Weijun revealed, “For the same mammography image, the discrepancy in annotations between an experienced senior radiologist and a junior physician can reach 30% or even higher. Therefore, there is enormous potential for the application of artificial intelligence in this field.”

The difficulty of developing AI for mammography is beyond the reach of ordinary companies.

“From the very inception of our R&D efforts, we have adhered to the latest ACR guidelines, while also referencing the NCCN guidelines, ACS guidelines, and the most recent Chinese Expert Consensus on the Diagnosis and Treatment of Breast Cancer. Both R&D and engineering personnel were required to start from scratch in learning about breast imaging and breast cancer, working closely with clinicians over an extended period to gain a deep understanding of clinical workflows, AI application scenarios, and physicians’ pain points,” stated Lin Qiang, Medical Product Director at Yitu Healthcare.

Clinically annotated data, professionally labeled, is not only key to building AI models but also the cornerstone of this breast AI.

“This AI aggregates tens of thousands of breast cancer imaging cases from multiple top-tier tertiary hospitals across China, featuring comprehensive patient positioning, advanced equipment, professional acquisition protocols, and high-clarity images, making it arguably the premier breast tumor database in China currently,” said Lin Qiang with considerable confidence. “The annotation process was meticulously designed, covering both hardware and software aspects.”

To ensure a more comprehensive description of lesion features, the R&D team immersed themselves in clinical frontline practice, consulting experts and reviewing guidelines to develop detailed annotation rules for each feature description.

To provide physicians with clearer visualization, the R&D team has equipped them with professional 5MP medical-grade monitors, ensuring sharper image display and reduced visual strain.

To alleviate the workload burden on annotating physicians and avoid prioritizing annotation speed alone, the R&D team recruited dozens of professionally qualified physicians who had undergone rigorous assessments to form an annotation team, thereby distributing the workload and preventing rushed annotations.

To ensure annotation quality, each mammogram was annotated by at least five physicians. Only annotations with high inter-rater agreement were accepted. Discrepant annotations were adjudicated by senior physicians, submitted to the team for review and voting, and finally determined by authoritative experts.

To enable more efficient oversight of the annotation workflow, the R&D team even developed a dedicated annotation management system prior to building the AI model.

The high costs of data curation and correlation, the cumbersome annotation process, and the numerous contentious labeling points once pushed the R&D team to the brink of collapse. Yet, the final model results did not disappoint these efforts and expectations; during its implementation in hospitals, the system has consistently received high praise from experts.

“In Western countries, it is already impressive for a radiologist to interpret ten mammograms a day. In China, however, that number is at least 50. For associate professors responsible for reviewing and signing off on reports, interpreting 100 or even 150 mammograms per day is commonplace. Under tight deadlines and heavy workloads, with no room for missing any lesions, physicians endure prolonged physical and mental stress,” revealed Professor Peng Weijun. “Artificial intelligence systems can significantly improve the speed and accuracy of lesion detection, reduce false positives and missed diagnoses, enhance reading consistency, and free physicians from burdensome repetitive tasks, allowing them to focus on truly innovative work.”

With the increase in international academic exchanges and the return of a growing number of top AI experts and scholars to China, it is not uncommon in the industry to leverage overseas public datasets for medical AI research and development. Objectively speaking, the emergence of public datasets and generalizable AI models has significantly propelled the development of medical artificial intelligence in China. However, in the R&D of mammography AI, models developed based on overseas public datasets have suffered a major setback.

Lin Qiang revealed that, unlike the predominantly fatty breast tissue common in Western women, Chinese women typically have dense breast tissue, where glandular obscuration and structural noise are more pronounced. This results in lower contrast between normal breast tissue and lesions, thereby imposing higher performance requirements on AI systems.

“AI for mammography developed based on public datasets may achieve a sensitivity of 95% or even over 99% in laboratory settings, but once deployed in clinical practice, its sensitivity drops significantly. It requires prolonged tuning and extensive data feeding, which inadvertently increases the burden on clinicians,” said Lin Qiang. “Furthermore, public datasets also suffer from issues such as low image quality, poor annotation quality, and inconsistent annotation standards.”

Taking a single mediolateral oblique (MLO) view from mammography as an example, Lin Qiang noted that a standard mammogram typically has a resolution as high as 4000x4000 pixels, totaling over 16 million pixels, resulting in large file sizes. However, to facilitate publication, public datasets often compress these images into standard JPEG format, leading to significant pixel loss. This substantially diminishes the representation of lesions, with microlesions potentially disappearing entirely. Consequently, the performance of AI models trained on such data is inevitably compromised.

Therefore, the advantages of AI developed based on real breast imaging data from Chinese women are clearly evident.

“We have specifically optimized the image-reading algorithms for this AI system, enabling direct ingestion and second-level processing of 16-megapixel images without any lag or crashes. This ensures that even the most minute lesions are detected with exceptional clarity, preserving all details and accurately reconstructing lesion morphology. In this regard, AI far surpasses human capability,” said Lin Qiang.

Currently, the vast majority of mammography services in China are concentrated in large and medium-sized cities, while nearly 800 million urban and rural residents served by primary healthcare institutions receive medical care that is significantly inferior to that available in urban areas; mammography is no exception.

“In the future, as breast AI becomes more mature, the cost of universal early screening for breast cancer will be significantly reduced. The approach to early screening for breast cancer will also evolve, gradually expanding from high-risk populations to all age-eligible women, thereby greatly increasing the likelihood of early detection and reducing overall societal healthcare expenditures,” stated Lin Qiang. “Meanwhile, by transforming expert-level diagnostic and treatment capabilities into accessible tools and empowering primary healthcare with AI, we can help alleviate the current shortage of radiologists in primary care institutions and enhance their capacity for early screening of breast diseases.”