Presenter Dr. Natally Horvat, PhD, and senior author Dr. Iva Petkovska from Memorial Sloan Kettering Cancer Center and colleagues trained two radiomics classifiers to predict pathological complete response using restaging MRI scans of 114 patients with locally advanced rectal cancers. All patients had received neoadjuvant therapy followed by total mesorectal excision at their institution between March 2012 and February 2016.
They then validated the models on an external dataset of 50 consecutive patients from a second institution. For these cases, two radiologists also evaluated the restaging MRI exams and classified the patients as either demonstrating a radiological complete response or radiological partial response.
Several models were developed and evaluated:
- Model A (radiomics analysis of 33 texture features on MRI)
- Model B (radiomics analysis of 91 texture, shape, and edge features on MRI)
- A combined model including models A and B, as well as the first radiologist's qualitative assessment
- A combined model including models A, B, and the second radiologist's qualitative assessment
Model B yielded an area under the curve (AUC) of 0.83 for predicting treatment response, and model A provided similar discriminative ability (p = 0.3). What's more, the researchers found that the combination of radiomics with radiologist interpretations led to better assessments of pathological complete response than the radiologists produced on their own.
"A radiomics classifier constructed on a single-center dataset showed good discriminative ability to predict [pathologic complete response] on an external dataset," the authors wrote. "The combined radiomics and radiologist model improved reliability of [pathologic complete response] prediction."
What else did they find? Stop by this presentation on Sunday to get all of the details.
