Robust Pancreatic Ductal Adenocarcinoma Segmentation with Multi-Institutional Multi-Phase Partially-Annotated CT Scans

TL;DR

This paper introduces a self-learning framework for pancreatic ductal adenocarcinoma segmentation in CT scans, leveraging a large dataset with limited annotations to improve accuracy and achieve performance comparable to radiologists.

Contribution

The study presents a novel semi-supervised approach that combines multiple teacher models and vessel information to effectively utilize unannotated images for PDAC segmentation.

Findings

Achieved 6.3% Dice score improvement over baseline

Attained Dice score of 0.71, comparable to radiologist variability

Utilized a large dataset of ~1,000 patients with mixed annotations

Abstract

Accurate and automated tumor segmentation is highly desired since it has the great potential to increase the efficiency and reproducibility of computing more complete tumor measurements and imaging biomarkers, comparing to (often partial) human measurements. This is probably the only viable means to enable the large-scale clinical oncology patient studies that utilize medical imaging. Deep learning approaches have shown robust segmentation performances for certain types of tumors, e.g., brain tumors in MRI imaging, when a training dataset with plenty of pixel-level fully-annotated tumor images is available. However, more than often, we are facing the challenge that only (very) limited annotations are feasible to acquire, especially for hard tumors. Pancreatic ductal adenocarcinoma (PDAC) segmentation is one of the most challenging tumor segmentation tasks, yet critically important for clinical needs. Previous work on PDAC segmentation is limited to the moderate amounts of annotated patient images (n<300) from venous or venous+arterial phase CT scans. Based on a new self-learning framework, we propose to train the PDAC segmentation model using a much larger quantity of patients (n~=1,000), with a mix of annotated and un-annotated venous or multi-phase CT images. Pseudo annotations are generated by combining two teacher models with different PDAC segmentation specialties on unannotated images, and can be further refined by a teaching assistant model that identifies associated vessels around the pancreas. A student model is trained on both manual and pseudo annotated multi-phase images. Experiment results show that our proposed method provides an absolute improvement of 6.3% Dice score over the strong baseline of nnUNet trained on annotated images, achieving the performance (Dice = 0.71) similar to the inter-observer variability between radiologists.