Amsterdam UMC, location VUmc
Anatomy & Neurosciences

Author Of 1 Presentation

Imaging Poster Presentation

P0539 - Artificial double inversion recovery images for cortical lesion visualization in multiple sclerosis (ID 817)

Speakers
Presentation Number
P0539
Presentation Topic
Imaging

Abstract

Background

Cortical lesions (CLs) in multiple sclerosis (MS) are clinically important, but highly inconspicuous on conventional clinical MRI. Double inversion recovery (DIR) is sensitive to CL detection, but difficult to implement in clinical practice and research settings, as it is difficult to set up and proper acquisition may take significant time due to the required inversion times (i.e.,~8 to 10 minutes). This work examines whether artificial intelligence can mitigate this dilemma through generation of artificial DIR images from –readily available– conventional clinical MR sequences.

Objectives

To determine whether artificially generated DIR (aDIR) images can be used for CL detection in MS and assess how this compares to conventionally acquired DIR (cDIR) images.

Methods

In this retrospective study, aDIR images were generated from conventional 1.5 Tesla 3D-T1 and 2D-proton density/T2 images in 73 patients with MS (49 RRMS, 20 SPMS, 4 PPMS) and 42 controls. A fully convolutional 3D conditional adversarial network following an adapted U-Net design with skip-connections was trained, using images of 58 patients and 34 controls. The remaining subjects were assigned to the test set for which artificial 3D-DIR images were generated. To determine detection reliability, precision and recall, the aDIR and cDIR images of subjects in the test set were blindly scored for CLs.

Results

A total of 626 CLs were detected on 15 aDIR images versus 696 on cDIR images (ICC=0.92, 95% confidence interval 0.68-0.98 (F(32.755)). Compared to cDIR images, CLs were detected on aDIR images with an average precision and recall of 0.84±0.06 and 0.76±0.09, respectively. The largest difference in CL discernibility was observed in frontal and temporal regions.

Conclusions

Artificially generated DIR images showed excellent reliability, precision and recall in detected cortical lesions when compared to conventionally acquired DIR images. The technique has the potential to broaden DIR availability and to enable retrospective implementation of cortical lesion detection with DIR. Histopathological and multi-center validation are necessary to formally compare sensitivity and specificity and cross-scanner robustness.

Collapse