Development and external evaluation of MRI and clinical models for pathological complete response in breast cancer
This study found that adding fixed pretreatment MRI signal summaries to clinical models did not significantly improve the external prediction of pathological complete response in breast cancer, suggesting that further assessment of calibration, uncertainty, and clinical utility is required before such MRI data can be used for treatment-response communication.
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
Before a woman undergoes surgery for breast cancer, doctors often administer powerful drugs to shrink the tumor first. This approach, known as neoadjuvant therapy, can sometimes eliminate the cancer entirely, a result doctors call a "pathological complete response." When this happens, the tumor disappears so completely that a pathologist cannot find any remaining cancer cells under a microscope. Achieving this outcome is a major victory, as it is linked to better long-term survival, especially for aggressive types of the disease. However, predicting who will achieve this total disappearance before treatment even begins is difficult. Currently, doctors rely on basic information like the patient's age and the specific biological markers of the tumor to make these estimates. The question has long been whether adding a detailed look at the tumor using magnetic resonance imaging (MRI) could make these predictions significantly more accurate, allowing for more personalized conversations between doctors and patients about what to expect.
A researcher at King's College London set out to answer this question by building and testing computer models designed to predict this outcome. The study did not involve new patients or new scans; instead, it used existing, anonymized data from two large, well-known breast cancer trials. The researcher first built the prediction tools using data from nearly one thousand patients in the I-SPY2 trial. These models were fed two types of information. The first was a standard clinical profile: the patient's age and the status of three key receptors on the cancer cells, which determine how the tumor behaves and responds to treatment. The second type of information came from the MRI scans taken before treatment started. Specifically, the model looked at twelve specific measurements derived from the brightness and contrast patterns of the tumor in the scan, summarizing the tumor's appearance without measuring its exact size or shape.
The goal was to see if adding these twelve MRI measurements to the basic age and receptor information would improve the model's ability to guess the outcome. To test this fairly, the researcher took the models that were built and "froze" them, meaning they were locked in place and could not be changed. These frozen models were then applied to a completely separate group of 166 patients from the older I-SPY1 trial, a group the models had never seen before. This step is crucial because it tests whether the model works in a different setting or if it simply memorized the patterns of the first group. The researcher compared how well the model using only age and receptors performed against the model that also included the MRI data.
The results showed that adding the MRI data did not improve the prediction. In fact, the model using only the basic clinical information performed slightly better than the one that included the MRI scans when tested on the new group of patients. The model with the MRI data did not provide a more accurate estimate of the chance of a complete response. Furthermore, the study looked at how stable these predictions were. If the researcher had built the model again using a slightly different set of patients from the same pool, would the predictions for an individual patient change drastically? The analysis revealed that the predictions from the model including MRI data were much less stable. The estimated probability of a cure for a specific patient could swing by nearly nine percentage points just by changing the group of people used to build the model, whereas the simpler model was much more consistent.
The study also examined how well the models calibrated their predictions. A good model should not only rank patients correctly but also predict probabilities that match reality. For instance, if a model predicts that 30 percent of patients will have a complete response, then roughly 30 percent of them should actually achieve it. In this study, all the models tended to overestimate the chances of success, predicting a higher likelihood of a cure than what was actually observed in the patients. The model that included MRI data was particularly prone to this overestimation. The researchers concluded that while MRI scans contain a vast amount of information, the specific way these twelve measurements were summarized and added to the basic clinical data did not help. The study explicitly ruled out the idea that this specific, compact MRI summary offers a reliable improvement for predicting treatment response in this context.
Ultimately, the research suggests that for the purpose of telling a patient their specific odds of a complete cure before treatment starts, adding these fixed MRI summaries to age and receptor status does not provide a clear benefit. The study found no evidence that the MRI data improved the accuracy of the predictions or that it made the models more reliable. The researchers emphasized that before such tools could be used to guide difficult medical decisions or conversations, they would need to demonstrate better accuracy in external tests and show that their predictions remain stable. Until then, the added complexity of the MRI data does not justify its use over the simpler, more consistent clinical information for this specific prediction task.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.