A pre-operative model that extracts features from medical images significantly improves predictions of survival for patients with pancreatic cancer compared with currently used clinical methods, according to research presented on Wednesday.
Healy
Patients with pancreatic cancer face poor odds, with a five-year survival rate of less than 10% internationally. Predicting patient survival at the beginning of treatment — a concept known as prognostication — has been investigated as way to improve outcomes for patients.
“This information can help decide what treatment is most appropriate for the patient; in particular, the question as to whether patients should have neoadjuvant chemotherapy prior to surgery,” said presenter and study author Gerard Healy, BM BCh, from Toronto General Hospital.
For the new study, Dr. Healy and colleagues used radiomics, the extraction of features from medical images, to develop a prognostic model for pancreatic cancer based on pre-operative CT. Radiomics has shown potential value for prognostication, but previous studies have lacked external validation.
The researchers studied the radiomics model in 352 pancreatic cancer patients who underwent pre-operative contrast-enhanced CT without neoadjuvant therapy at five North American hospitals. The patients subsequently underwent surgery at the University Health Network in Toronto.
The researchers then externally validated the radiomics model on 215 patients who underwent surgery at St. Vincent’s University Hospital in Dublin following pre-operative CTs performed at 34 hospitals.
“We are the first group in pancreas radiomics to conduct such a robust external validation of our model,” Dr. Healy said.
The radiomics score significantly improved prognostication for patients compared to using clinical information alone. In the external validation group, the radiomics score was the dominant predictor of overall and disease-free survival. No other clinical features were significantly associated with overall and disease-free survival. Radiomics performed similarly to pathological information, which is only available post-operatively.
“Our results found that the model performed better than the clinical prognostic variables currently available in clinical practice,” Dr. Healy said.
AI Methods May Be More Effective
The model could help guide pre-operative therapy decisions, but more improvement is needed before it reaches that level. Dr. Healy noted that prior studies in this field by other groups have had positive results, but they used smaller study populations and included minimal to no external validation.
“Since we are the first group to perform such robust external validation, our conclusion is that radiomics is not currently ready for clinical implementation in pancreatic cancer prognostication,” Dr. Healy said. “We therefore conclude that the results of those prior studies were overly optimistic.”
When asked at the session if any particular tumor feature on CT stood out for predictive value, Healy pointed to tumor attenuation.
“Low attenuation tumors do worse,” he said.
Dr. Healy believes that the future of prognostication in pancreatic cancer lies in more advanced artificial intelligence methods like deep learning. His lab is shifting its focus to that through continuing collaboration between the St. Vincent’s University Hospital in Dublin and the lab of radiologist Masoom Haider, MD, at the University of Toronto.
Dr. Healy’s two-year clinical-research fellowship at the University of Toronto was made possible through a scholarship from the Faculty of Radiologists, Royal College of Surgeons in Ireland.
Access the presentation, “Pre-operative Radiomics Model For Prognostication In Resectable Pancreatic Adenocarcinoma: Multi-institutional Development And External Validation,” (SSGI11) on demand at Meeting.RSNA.org.
