Posted: Tuesday, January 17, 2023
Machine learning algorithms trained on whole-slide images may be useful in predicting outcomes from radical prostatectomy in patients with prostate cancer, according to an article published in JCO Clinical Cancer Informatics. The investigators reported that the novel algorithm performed similarly to standard clinical multivariable risk assessment tools using data from whole-slide images alone.
“Machine learning approaches to pathologic diagnosis and prognosis are a promising and increasingly used strategy,” said study author Nathan Paulson, MD, of Stanford University, and colleagues. “Through models such as this, implementation of larger, more complex risk stratification approaches may improve care in prostate cancer.”
The investigators retrospectively reviewed 361 whole-slide images from 107 patients with grade group 2 or 3 prostate cancer. The machine learning pipeline proceeded as follows: Images were preprocessed for use by a machine learning algorithm, relevant features were extracted, and the algorithm produced an outcome prediction. The authors noted that a pretrained neural network was used for feature extraction, and the extreme gradient boosting classifier was used for binary outcome prediction.
The group used the area under the receiver operating characteristic curves to analyze the performance of the model. When predicting outcomes of the entire cohort, the model achieved a score of 0.72 (95% confidence interval [CI] = 0.62–0.81), which the authors noted is similar to that achieved with clinical risk assessment tools. The authors also explored how different disease states performed independently, finding that the model performed better when predicting outcomes from patients with grade 3 disease (0.89, 95% CI = 0.79–1.00) and worse when modeling outcomes from patients with grade 2 disease (0.65, 95% CI = 0.53–-0.79). Lastly, the authors noted that in future studies, the use of clinical information along with whole-slide imaging has the potential to improve the accuracy of the model.
Disclosure: For full disclosures of the study authors, visit ascopubs.org.