When surgeons remove pancreatic cancer via pancreaticoduodenectomy, the goal is an R0 resection — removing the tumor with at least 1mm of cancer-free tissue on all sides. An R1 resection means cancer cells are within 1mm of the cut edge, indicating incomplete removal.
The difference matters enormously: R0 resections are associated with significantly better long-term survival, while R1 resections carry higher risk of local recurrence. Currently this determination can only be made after surgery by examining the specimen.
A preoperative CT-based prediction tool could change surgical planning and patient counseling, enabling better decisions about surgical approach or whether neoadjuvant chemotherapy before the operation might improve resectability.
Researchers analyzed preoperative CT scans from 86 patients with pancreatic head adenocarcinoma who underwent pancreaticoduodenectomy — 34 with R0 and 52 with R1 outcomes — giving 258 images total (three slices per patient).
The radiomics pipeline outlined tumor regions of interest, fitted them to rectangular regions using mathematical equations, then enhanced textures with wavelet transforms and fractional differential operators to amplify subtle patterns before feature extraction.
Principal component analysis reduced features to the most informative ones, and a support vector machine classifier distinguished R0 from R1 cases. Leave-one-out cross-validation assessed performance while minimizing overfitting.
The radiomics model achieved an accuracy of 84.88% and an AUC of 0.8614 in distinguishing R0 from R1 resections based on preoperative CT alone — meaningful performance for a preliminary study.
Statistical analysis identified two texture features from wavelet-transformed images as significantly different between R0 and R1 cases, reflecting how pixel intensities are distributed across the tumor in ways that correlate with margin status.
These results suggest CT images contain information about tumor-tissue relationships near the surgical margin that human readers cannot reliably perceive, but that AI texture analysis can detect.
A key innovation was using fractional differential operators to enhance CT textures before feature extraction. Unlike standard derivatives that highlight sharp edges, fractional derivatives extract subtle mid-range patterns preserving information standard processing might discard.
Wavelet transforms decomposed images into multiple frequency bands, capturing both coarse and fine structural information simultaneously — picking up patterns at different spatial scales within the tumor.
Combining these enhancement techniques with statistical texture features creates a rich fingerprint that reflects biological properties — like cellular density near the surgical margin — that pathologists assess post-operatively.
This preliminary study demonstrates the potential of radiomics to provide preoperative information about surgical resection margins in pancreatic head cancer, supporting the hypothesis that CT images contain extractable biological information.
Validation in larger multi-center cohorts is needed before clinical adoption. Automated segmentation and deeper learning architectures may further improve performance.
If validated, a reliable preoperative radiomics-based margin prediction tool could help surgeons identify patients likely to benefit from aggressive approaches, or those who should receive neoadjuvant therapy to improve resectability before surgery.