Ultrasound-based deep learning-derived radiomics for preoperative prediction of sentinel lymph node metastasis in invasive breast cancer: a two-center retrospective study

In the journey through breast cancer treatment, one of the most critical decisions happens before the first incision is made: determining whether the cancer has spread to the nearby lymph nodes. These small, bean-shaped organs act as the body's first line of defense, filtering fluid from the breast. If cancer cells have traveled there, the surgical plan changes significantly. Traditionally, doctors must perform a sentinel lymph node biopsy, a procedure where a surgeon removes a specific node to check for cancer under a microscope. While effective, this is an invasive step that carries risks of swelling and discomfort, and it is not always perfect; sometimes, the test misses the cancer, and other times, it leads to unnecessary surgery for patients whose cancer has not spread. For decades, researchers have searched for a way to predict this spread using non-invasive tools, hoping to spare patients from procedures they do not need. Ultrasound, a common imaging technique that uses sound waves to create pictures of the body, has long been a standard tool for looking at breast lumps, but its ability to see the invisible spread of cancer cells has been limited by the human eye's ability to interpret the images.

A team of researchers from two hospitals in China has taken a new approach to this old problem by teaching a computer to see patterns in ultrasound images that the human eye might miss. They focused on a specific type of breast cancer called invasive breast cancer, where the tumor has the potential to break through the breast tissue and travel to the lymph nodes. The team gathered ultrasound images of tumors from 246 patients, all of whom had undergone surgery to confirm whether their lymph nodes contained cancer. They split these patients into two groups: one group from a hospital in Guangzhou to teach the computer, and a separate group from a different hospital to test if the computer could apply what it learned to new, unseen patients. The computer did not look at the images the way a radiologist does, by noting the shape or brightness of the tumor. Instead, it used a deep learning system, a type of artificial intelligence, to break the images down into thousands of tiny mathematical details. The system then used a decision-making algorithm to find the specific combination of these details that best predicted whether the cancer had reached the lymph nodes.

The results of this study show that the computer model achieved a moderate level of success in making these predictions. When tested on the group of patients used to train the system, the model achieved an AUC of 0.73. When the researchers tested the same model on the independent group of patients from the second hospital, the performance remained consistent, with the model achieving an AUC of 0.74. This consistency across two different hospitals, which used different ultrasound machines, suggests that the computer learned a genuine pattern rather than just memorizing the specific images it was shown. However, the model is not a perfect crystal ball. In the test group from the second hospital, the model was very good at correctly identifying patients who did not have cancer spread, with a success rate of about 84 percent. This is a valuable finding because it means the tool could potentially help doctors feel more confident in skipping the invasive biopsy for certain low-risk patients.

Yet, the study also highlights where the tool falls short. The model was less successful at catching every single case where cancer had spread, correctly identifying those cases only about 65 percent of the time. Because it missed some cases of spread, the researchers emphasize that this tool cannot replace the standard biopsy procedure on its own. If a doctor relied solely on this computer prediction to rule out cancer spread, they might mistakenly send some patients home without the necessary treatment. The authors describe the model as a helpful assistant rather than a replacement, a tool that could add a layer of information to the decision-making process. They note that the study was retrospective, meaning it looked back at data that had already been collected, and the number of patients was relatively small. The researchers also point out that the computer was not taught to explain why it made a certain decision, which is a common challenge with this type of advanced artificial intelligence.

The path forward, according to the authors, involves testing this tool on much larger groups of patients in the future. They suggest that before this technology can be used in everyday clinics, it needs to be proven in prospective studies, where patients are followed forward in time to see if the predictions hold up in real-world practice. The goal is to see if combining this computer analysis with other standard information, like the size of the tumor or the patient's age, can improve the accuracy even further. For now, the study offers a promising glimpse into a future where artificial intelligence might help doctors make more personalized decisions, potentially sparing some women from unnecessary surgery while ensuring that those who need it receive it. The work demonstrates that while the technology is not ready for prime time, it has moved past the theoretical stage and shown real, measurable potential in a clinical setting.