A Small-Lesion-Focused Framework for Early Detection of Sub-2 Cm Pancreatic Tumours with Synthetic Data Enhancement

Abstract

Pancreatic ductal adenocarcinoma (PDAC) remains among the most lethal solid-organ malignancies, with five-year survival rates that lag far behind most other cancers primarily because the disease is usually diagnosed after it has already advanced beyond curative resection. A growing body of machine learning and deep learning research has targeted automated pancreatic cancer detection on CT and MRI, spanning classical texture-based pipelines (Discrete Wavelet Transform denoising, Fuzzy C-Means segmentation, Gray-Level Co-occurrence Matrix feature extraction) and modern deep architectures (U-Net, Swin-UNet, Vision Transformers, Graph-CNN hybrids, DenseASPP). Reported accuracy in this literature is consistently strong, typically in the 85-96% range, but these figures are computed across a case mix dominated by tumours already conspicuous on imaging. Tumours under 2cm, precisely the population in which earlier detection would translate into the largest survival benefit, are systematically underrepresented in how performance is measured and reported; a statistic repeatedly cited in this literature holds that routine abdominal CT misses on the order of 40% of pancreatic tumours below this size threshold. This paper (1) synthesizes the existing detection, segmentation, and classification literature specifically through the lens of the small-lesion problem, (2) formalizes the small-lesion evaluation gap as a distinct, underserved research question rather than an incidental limitation, (3) proposes a four-stage detection framework and a stratified evaluation protocol that reports sensitivity separately for sub-2cm and larger tumors, and (4) demonstrates the pipeline mechanics and evaluation protocol on a synthetic proof-of-concept dataset of 2,000 simulated patients, where three classifiers reach 92.6-93.8% sensitivity on simulated sub-2cm malignant cases against a simulated routine-CT-read baseline of 55.6%. These synthetic results are offered strictly as a demonstration that the pipeline and stratified-reporting protocol function as intended; validation on real, ethically sourced imaging data is proposed as the required next phase before any clinical interpretation is warranted.