Cascade Deep Learning (nnU-Net + ResNet-18) for Meniscus Segmentation and Three-Grade Injury Classification on Knee MRI: A Single-Centre Study with Reader Assistance Evaluation

The human knee is a marvel of engineering, relying on two C-shaped pads of cartilage called menisci to cushion the joint and keep it stable. When these pads tear or wear down, the result is pain, swelling, and a loss of mobility that often requires surgery. To diagnose these injuries, doctors rely on magnetic resonance imaging, or MRI, which creates detailed cross-sectional pictures of the inside of the knee. However, reading these images is a slow and demanding task. A radiologist must scan through hundreds of thin slices, looking for subtle changes in the texture of the cartilage that signal a tear. This process is not only time-consuming but also prone to human error; two different doctors might look at the same scan and disagree on whether a patient has a minor wear-and-tear issue or a severe tear that needs immediate repair. This inconsistency can delay treatment or lead to unnecessary procedures.

In an effort to make this process faster and more reliable, researchers have begun teaching computers to read these scans. This field, known as deep learning, involves training computer programs to recognize patterns in images much like a human brain learns to recognize faces. The goal is not to replace the doctor, but to create a tool that acts as a second pair of eyes, highlighting areas of concern and offering a second opinion. A new study from a hospital in China explores how well such a system works specifically for meniscus injuries. The researchers built a two-step computer system that first finds the meniscus in the MRI scan and then analyzes it to determine if it is healthy, degenerating, or torn. They tested this system on hundreds of patients and then watched how it affected the performance of junior doctors, aiming to see if the computer could help less experienced clinicians make better and faster diagnoses.

The researchers developed a computer framework that works in a specific sequence, much like a team where one person locates an object and another person inspects it. First, the system uses a sophisticated algorithm to scan the MRI images and draw a precise outline around the meniscus, separating it from the surrounding bone and fluid. This step is crucial because it tells the computer exactly where to look, ignoring the rest of the knee. Once the meniscus is isolated, a second part of the system examines the isolated image to classify the injury. It sorts the condition into one of three categories: healthy, degenerative (showing signs of wear), or a grade-III tear (a full tear that often requires surgery). The system was trained on data from 608 knee scans collected over four years, all taken on the same type of 1.5-Tesla MRI machine. The images included both front-to-back and side-to-side views, allowing the computer to see the meniscus from multiple angles.

When the researchers tested the system on its own, the results were mixed but promising. The computer was very good at finding the meniscus when it was healthy, successfully outlining it in most cases. However, when the meniscus was injured, the computer struggled to draw the outline as precisely, likely because tears change the shape and texture of the tissue, making it harder to distinguish from the background. Despite this difficulty in outlining, the system's ability to diagnose the injury was decent. When combining the views from both the front and side angles, the computer correctly identified whether a meniscus was torn or not about 74 percent of the time. It was better at spotting tears when looking at the front view of the knee compared to the side view. While the system could not perfectly distinguish between mild wear and a full tear on its own, it showed a clear ability to flag serious injuries.

The most significant part of the study, however, was not just how the computer performed alone, but how it helped human doctors. The researchers recruited two junior orthopedic residents, each with less than three years of experience in reading musculoskeletal images, to diagnose a set of 100 knee cases. First, the doctors looked at the scans without any help. Later, after a two-week break to ensure they did not remember the cases, they looked at the same scans again, this time with the computer's assistance. The computer provided them with a probability score indicating how likely a tear was, along with a visual map that highlighted the specific areas of the meniscus the computer found suspicious.

The results of this human-machine partnership were striking. When the junior doctors worked alone, they achieved an AUC of 0.64. When they used the computer's guidance, their performance improved to an AUC of 0.78. More importantly, the computer helped them catch more severe tears. Without help, the junior doctors missed 40 percent of the grade-III tears; with the computer's assistance, they caught 80 percent of them. The tool also made the doctors faster. The average time it took to diagnose a single case dropped from nearly 45 seconds to just under 28 seconds, a reduction of 38 percent. This speedup suggests that the computer helped the doctors focus their attention, reducing the time spent searching for the injury. Additionally, the agreement between the two doctors improved when they both used the computer, suggesting that the tool provided a consistent standard that helped reduce their individual differences in judgment.

The study concludes that while this computer system is not yet perfect enough to replace a specialist, it serves as a powerful assistant. It acts as a safety net, particularly for less experienced clinicians, helping them spot serious injuries they might otherwise miss and doing so more quickly. The researchers noted that the system's performance was limited by the difficulty of defining the exact boundaries of torn tissue and by the subtle visual similarities between degeneration and early tears. They also pointed out that their study was conducted at a single hospital, meaning the system needs to be tested on patients from different locations and with different types of MRI machines before it can be widely adopted. Nevertheless, the findings demonstrate that a well-designed computer tool can effectively support human decision-making, turning a slow and variable process into a faster, more reliable one.