Comparison of 2-Dimensional and 3-Dimensional Image Segmentation on Gastrointestinal Bleeding Detection in DSA using Convolutional Neural Networks

Sunday, March 5, 2023

3:18 PM – 3:27 PM

Location: Phoenix Convention Center, 224AB

Presenting Author(s)

Derek Smetanick, Student

Student
Arizona State University

Disclosure(s): No financial relationships to disclose

Author/Co-author(s)

SN

Sailendra Naidu, MD

Interventional Radiologist
Mayo Clinic Arizona
GK

Grace Knuttinen, MD, PhD

Interventional Radiologist
Mayo Clinic Arizona
JK

J. Scott Kriegshauser, MD

Interventional Radiologist
Mayo Clinic Arizona
Rahmi Oklu, MD, PhD (he/him/his)

Professor
Mayo Clinic Arizona
Indravadan J. Patel, MD, FSIR

Division Chief - Interventional Radiologist
Mayo Clinic Arizona
AW

Alex Wallace, MD

Interventional Radiologist
Mayo Clinic Arizona
Sadeer Alzubaidi, MD

Interventional Radiologist
Mayo Clinic

Disclosure(s):

Derek Smetanick, Student: No financial relationships to disclose

Rahmi Oklu, MD, PhD: Boston Scientific: Advisory Committee or Review Panel Member (Ongoing), Consultant (Ongoing); InTumo Therapeutics, Inc: Board Membership (Ongoing), Intellectual Property Rights (Ongoing), Management Position (Ongoing), Ownership Interest (Ongoing); Obsidio, Inc: Board Membership (Terminated, August 1, 2022), Intellectual Property Rights (Terminated, August 1, 2022), Management Position (Terminated, August 1, 2022), Ownership Interest (Terminated, August 1, 2022)

Indravadan J. Patel, MD, FSIR: No financial relationships to disclose

Purpose:

To train a convolutional neural network that locates gastrointestinal bleeding in digital subtracted angiography (DSA) prior to transarterial embolization

Materials and Methods:

All mesenteric artery angiograms and arterial embolization DSA images from 2007 – 2021 taken in the interventional radiology department at Mayo Clinic Phoenix were analyzed. Derived from 26 patients, 39 unique series of bleeding were augmented to train 2D and 3D convolutional neural networks for image segmentation. Using the same images and augmentation methods, the 2D network was trained on 3,548 images and tested on 394 images and the 3D network was trained on 316 3D objects and tested on 35. Both 128x128 and 256x256 pixel sizes were used to train separate 2D convolutional neural networks. For each case, cropped images focused on the GI bleed and uncropped images were examined. Two different post image processing techniques were used for each case including merging all masks within a series or analyzing each individual mask. In total, there were 10 different test cases.

Results:

Accuracy was defined as a correct identification of the bleeding either in partial or full. Accuracy with false positive was defined as accurate but had at least one pixel of bleed identified where it should not have. Since minor false positives were not concerning as long as the bleed was identified, both categories contributed to percent accurate. The accuracy is recorded in the table.

Conclusion:

While DSA offers a real-time approach of locating the site of the bleed, many sources of bleeding may not be easily visible on the angiogram. DSA is significantly less sensitive in detection of extravasation compared to computed tomography (CT) studies. Our study addresses this need for an artificial intelligence to locate the GI bleeding in DSA to assist interventional radiologists. The neural net demonstrated positive results with the accuracy breaking over 50% in identifying bleeding in uncropped images. Further studies need to be conducted before implementing in DSA for real time predictions.