Knowledge Distillation-Based Lightweight Generative Large Language Models for Standardized Quality Control of Chinese Radiology Reports
This paper presents a locally deployable, knowledge distillation-based 4B language model system that effectively automates the standardized quality control of Chinese radiology reports, achieving high precision and significantly improved inference speed compared to larger teacher models.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the quiet hum of a hospital, a radiologist's work often culminates in a single, critical document: the radiology report. This text, born from the images of X-rays, CT scans, and MRIs, serves as the bridge between a machine's observation and a doctor's decision. It tells the story of what is happening inside a patient's body, guiding treatments that can save lives. Yet, like any human endeavor, this writing is prone to slips. A missing detail, a contradictory sentence, or a non-standard description can obscure the truth, potentially leading to a misdiagnosis or a delayed intervention. For decades, the solution to these errors has been a second pair of human eyes, a manual check that is slow, exhausting, and inevitably inconsistent. In recent years, powerful artificial intelligence systems have emerged that can read and understand text with human-like fluency, offering a way to automate this quality check. However, these "large" models are often too massive to run inside a hospital's own computers, forcing institutions to send sensitive patient data to the cloud, which raises serious privacy concerns.
This tension between the need for powerful intelligence and the need for local privacy is the stage for a new study by researchers from Anhui University of Chinese Medicine and a local medical technology company. They set out to solve a specific problem: how to build a system that can check Chinese radiology reports for errors with the accuracy of a giant supercomputer, but small enough to run on a standard hospital server without ever sending data off-site. The researchers did not simply try to shrink a big model; instead, they used a technique called knowledge distillation. Imagine a master teacher who knows everything about a subject but is too slow to answer every question instantly. The researchers had this master teacher review thousands of reports and explain its reasoning. They then trained a much smaller, faster student model to learn from those explanations. The goal was to capture the master's wisdom in a compact form that could work instantly and privately within the hospital walls.
The team began by defining exactly what a "good" report looks like. They created a strict set of thirteen rules based on national medical standards, covering everything from logical consistency to specific formatting requirements. For instance, a report must not contradict itself, must describe the size of a lesion if it is measurable, and must order findings by their importance to the patient's health. To teach their system these rules, they first selected the best available "teacher" from three powerful, cloud-based artificial intelligence models. After testing them, they found that one model, which they identified as Doubao-seed1.6-non-thinking, was the most accurate at spotting errors, correctly identifying violations with a high degree of precision. This model became the expert that would guide the rest of the process.
With the expert chosen, the researchers faced a challenge: real-world reports that contained errors were rare compared to those that were perfect. To teach the student model effectively, they needed more examples of mistakes. They used the expert teacher to generate thousands of synthetic reports, carefully introducing specific errors into otherwise perfect text to create a balanced training set. This allowed the student model to see a wide variety of problems without waiting for them to happen naturally. They then selected three small, lightweight models to act as students. These models were designed to be fast and efficient, capable of running on local hardware. Using a method called knowledge distillation, they fed the student models the vast collection of reports and the expert teacher's corrections, allowing the small models to learn the patterns of a good report without needing to store the massive amount of data required by the big models.
The results showed that the approach worked remarkably well. The best student model, a compact system named Qwen3-4B-Instruct-2507, learned to identify errors with an average precision of 0.90, a score very close to the expert teacher it was trained by. More importantly, the speed difference was staggering. While the expert teacher took over eighteen seconds to analyze a single report, the lightweight student model completed the same task in just 0.41 seconds. This represents a speed increase of more than forty times, transforming a process that would have been a bottleneck into one that happens almost instantly. The system was then tested in a real clinical environment on the Anhui Province Medical Imaging Cloud Platform, where it processed nearly seventeen thousand real reports over the course of a week. In this live setting, the model flagged errors that led radiologists to revise 792 reports, and the model's alerts were accepted for 74 percent of those revisions, demonstrating that it could catch genuine issues in the flow of daily work.
The study confirms that it is possible to bring high-level artificial intelligence quality control directly into hospitals without compromising patient privacy or requiring expensive, specialized hardware. By distilling the knowledge of a massive cloud-based model into a small, local one, the researchers created a tool that is both powerful and practical. The system successfully identified logical inconsistencies, missing content, and formatting errors across different types of scans, from X-rays to MRIs. While the researchers noted that the model is not perfect and still struggles slightly with some complex, open-ended rules, it represents a significant step forward. It offers a way to standardize the quality of medical reports, reducing the risk of human error and allowing radiologists to focus more on diagnosis and less on repetitive checking. The work suggests that the future of medical text processing may not rely on sending data to the cloud, but on bringing the intelligence of the cloud down to the local server, making advanced quality control accessible to hospitals of all sizes.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.