Machine learning may guide use of neoadjuvant therapy for pancreatic cancer
Click Here to Manage Email Alerts
Key takeaways:
- A machine learning algorithm performed better than CT scans for predicting lymph node metastases and positive resection margins.
- Researchers plan a prospective trial to use the technology to help stratify patients for neoadjuvant treatment.
Machine learning algorithms can help predict positive resection margin and lymph node metastases among patients with pancreatic ductal adenocarcinoma, according to study results.
The approach yielded greater positive predictive values than CT scan for both variables, findings presented at Society of Surgical Oncology Annual Meeting showed.
“This hopefully can give providers the ability to identify patients with resectable pancreatic cancer who may benefit from neoadjuvant therapies,” researcher Jeremy Chang, MD, MS, surgery resident at University of Iowa Hospitals, said during a press conference.
Background and methodology
Pancreatic cancer is the third leading cause of cancer-related death, with a disproportionately high mortality rate compared with incidence due to most patients being diagnosed at advanced stages.
Approximately 15% to 20% of cases are deemed curable with surgery, according to study background. However, up to 80% of patients who undergo surgery develop local or distant recurrence, with key risk factors including lymph node metastasis, positive margins after surgery, larger tumor size and no receipt of chemotherapy.
“A recent novel notion is there may be patients with resectable tumors at time of diagnosis who would actually benefit from neoadjuvant therapy or chemoradiation before surgery,” Chang said. “The question now is, how do we find who those patients are?”
Chang and colleagues conducted a pilot study to assess the potential of machine learning — which uses algorithms to learn and recognize patterns from input data — to predict lymph node metastases or positive resection margins from preoperative scans.
Researchers used a 3-D convolutional neural network, optimized to process pixel or image data.
The network can be divided into three segments and 17 layers, Chang said. The first input layer consists of a CT image, followed by 12 layers of feature extraction, and then four layers of classification or output.
The cohort included adults diagnosed with pancreatic ductal adenocarcinoma who underwent pancreatectomy at University of Iowa Hospitals between 2015 and 2021. All patients had viable preoperative CT and postoperative pathology.
The analysis included 79 patients with a combined 480 CT images. The margin portion of the study also included 31 patients with unresectable locally advanced disease who served as positive controls.
Researchers divided patients into a training group — which allowed the algorithm to learn and develop its pattern of recognition — and a validation group.
Results
The lymph node status portion of the study included a training group of 59 patients with a combined 340 images, and a validation group of 20 patients with a combined 140 images.
Results of a per-patient analysis showed a sensitivity of 100% (95% CI, 80-100) and specificity of 60% (95% CI, 23-93).
Researchers reported a prediction accuracy of 90%, a positive predictive value of 88% (95% CI, 66-88) and a negative predictive value of 100% (95% CI, 44-100).
The margin status portion of the study included a training group of 83 patients with a combined 629 images, as well as a validation group of 27 patients with a combined 252 images.
Results showed a prediction accuracy of 81%, a positive predictive value of 80% (95% CI, 64-98) and a negative predictive value of 82% (95% CI, 59-94).
For context, the positive predictive value of CT scans — the most common modality for pancreatic cancer diagnosis and assessment — is 73% for identifying positive nodes and 68% for determining whether resection margins will be positive, Chang said.
“Future directions for this study will include increasing size of the training and testing cohorts to increase generalizability,” Chang said. “We’re also planning to use this technology to develop a prospective clinical trial to help stratify patients for neoadjuvant treatment.”
Perspective
Back to Top
Brett L. Ecker, MD
The abstract by Chang and colleagues highlights the novel applications of radiomics to the field of surgical oncology. Radiomics is an exciting methodology that, most briefly, converts medical images into big data. Vast amounts of qualitative imaging features can be abstracted from routine clinical imaging using data-characterizing algorithms. Some of these methods can be time- and expertise-limited, so this proof of concept was particularly exciting because no manual image segmentation was needed — potentially expanding its applicability.The big question is how these radiomic approaches will be incorporated into clinical practice. One possibility is that accurate identification of aggressive local disease — defined either by margin positivity at surgical resection, or regional lymphadenopathy — may steer clinicians towards neoadjuvant strategies prior to attempts at surgery. The assumption is that these high-risk pathologic features, which traditionally cannot be determined prior to surgery, are best initially managed with chemotherapy and/or chemoradiotherapy. Because nodal involvement and margin assessment are not available preoperatively, it’s not surprising that these variables are not included in the post-hoc Forest plots of the neoadjuvant trials for pancreas cancer (eg, PREOPANC). Thus, the translational significance of these findings ultimately will be defined by the prospective evaluation of these radiomic signatures to see if oncologic outcomes can be improved. The current wave of neoadjuvant trials in pancreas cancer are comparing sequencing strategies to better define the optimal timing of chemotherapy and/or chemoradiotherapy relative to surgical resection. Hopefully the incorporation of radiomic correlates will further clarify those patients most likely to benefit from each treatment sequence.
Collapse
Chang J, et al. Machine learning-based investigation of prognostic indicators for oncologic outcome of pancreatic ductal adenocarcinoma. Presented at: Society of Surgical Oncology Annual Meeting; March 22-25, 2023; Boston.
Click Here to Manage Email Alerts