---

About

Interfaces

This algorithm implements all of the following input-output combinations:

Challenge Performance

Date	Challenge	Phase	Rank
Nov. 28, 2024	PANORAMA	Development Phase - Sanity Check	1
Jan. 23, 2025	PANORAMA	Development Phase - Tuning	10
Feb. 23, 2025	PANORAMA	Testing Phase	3

Model Facts

Summary

We present an overview of our proposed method at the PANORAMA 2024 Grand Challenge, which aims to validate Artificial Intelligence algorithms for automated 3D detection and diagnosis of pancreatic ductal adenocarcinoma cancer (PDAC) from abdominal contrast-enhanced computed tomography. Our algorithm consists of four main sequential steps: (1) first, a low-resolution U-Net was trained to segment the pancreas parenchyma, (2) then a second full-resolution U-Net was trained to perform voxel-level segmentation encompassing all labels provided in the challenge, namely PDAC lesions, veins, arteries, pancreatic parenchyma, main pancreatic duct, and common bile duct. Third, (3) lesion risk score and main pancreatic duct maximum diameter were computed from the full-resolution U-Net. Lastly (4) the final probability of PDAC for a given pancreatic lesion was defined as a linear combination of the lesion risk and the main pancreatic duct maximum diameter.

Mechanism

First, a low-resolution 3D model segments the pancreatic parenchyma from the full contrast-enhanced computed tomography (CECT) scan and this predicted segmentation is used to crop the scan around a region of interest. Second, the cropped input is given to a full-resolution 3D model that segments the pancreatic parenchyma, the main pancreatic duct (MPD), the common bile duct, the veins, the arteries and any potential pancreatic ductal adenocarcinoma cancer (PDAC) lesion. A first lesion detection map is generated: the softmax is thresholded at 0.001, resulting in a binary map of connected components. Each component is assigned a lesion risk score based on the mean probability of its constituent voxels. Third, features are extracted: the lesion risk from the lesion detection map, and the maximum MPD diameter computed from the predicted segmentation. Finally, a logistic regression model, taking the lesion risk and the MPD diameter as inputs, updates the values of the detection map. The detection map is finally decropped to the initial input size.

Validation and Performance

Here are the following metrics obtained on the testing hidden cohort (see https://panorama.grand-challenge.org/evaluation/testing-phase/leaderboard/)

Metric	Value
AP	0.700447253707709
AUROC	0.9090422986395289
AP 95% (Upper)	0.7516689507715801
AP 95% (Lower)	0.647482265893604
Number of cases	957
AUROC 95% (Upper)	0.9287785129651259
Number of lesions	323
AUROC 95% (Lower)	0.8876036257598738
PICAI score (95% CI)	[0.76993676, 0.83799451]
Lesion TPR at FPR 0.1	0.7461300492286682
Lesion TPR at FPR 0.01	0.30030959844589233
Lesion TPR at FPR 0.001	0.11455108225345612

Uses and Directions

This algorithm was developed for research purposes only.

Warnings

Common Error Messages

Information on this algorithm has been provided by the Algorithm Editors, following the Model Facts labels guidelines from Sendak, M.P., Gao, M., Brajer, N. et al. Presenting machine learning model information to clinical end users with model facts labels. npj Digit. Med. 3, 41 (2020). 10.1038/s41746-020-0253-3