A large review of 57 studies involving nearly 12,000 participants looked at how machine learning can analyze medical images. The goal was to see if these computer models could better predict risks for patients with bladder cancer, such as whether a tumor has invaded nearby muscle or is high-grade.
The analysis found that radiomics-based machine learning performed well in several areas. For example, using MRI and CT scans combined with clinical data showed high accuracy in identifying muscle invasion and high-grade tumors. Different imaging types, including ultrasound, also showed some ability to detect these risks, though results varied by method.
While the technology shows promise for helping doctors understand cancer risk before surgery, there are important reasons to be cautious. The researchers noted that many of the original studies had flaws or a high risk of bias. Because of these quality issues, the findings are not yet ready to change how doctors treat patients in everyday clinical practice.
Common questions
How accurate is machine learning at finding muscle invasion?
The study found that radiomics-based machine learning showed high accuracy for identifying muscle invasion. For example, MRI-based methods reached an AUROC of 0.916, while CT-based methods reached 0.893. When combined with clinical features, the accuracy for MRI was even higher at 0.921.
Can these tools help identify high-grade tumors?
Yes, machine learning showed promise in diagnosing high-grade tumors. CT-based radiomics had an AUROC of 0.874, and MRI combined with clinical features reached 0.919. Ultrasound-based methods were also tested but showed a lower score of 0.750 for identifying high-grade tumors.
Is this technology ready to be used in hospitals today?
Not yet. While the results are interesting, the researchers noted that the evidence is not ready for clinical use. This is because many of the studies included in the review had methodological flaws and a high risk of bias, meaning more reliable research is needed first.
