Cerebrospinal fluid flow artifact reduction with deep learning to optimize the evaluation of spinal canal stenosis on spine MRI

PURPOSE

The aim of study was to employ the Cycle Generative Adversarial Network (CycleGAN) deep learning model to diminish the cerebrospinal fluid (CSF) flow artifacts in cervical spine MRI. We also evaluate the agreement in quantifying spinal canal stenosis.

METHODS

For training model, we collected 9633 axial MR image pairs from 399 subjects. Then, additional 104 image pairs from 19 subjects were gathered for the test set. The deep learning model was developed using CycleGAN to reduce CSF flow artifacts, where T2 TSE images served as input, and T2 FFE images, known for fewer CSF flow artifacts. Post training, CycleGAN-generated images were subjected to both quantitative and qualitative evaluations for CSF artifacts. For assessing the agreement of spinal canal stenosis, four raters utilized an additional 104 pairs of original and CycleGAN-generated images, with inter-rater agreement evaluated using a weighted kappa value.

RESULTS

CSF flow artifacts were reduced in the CycleGAN-generated images compared to the T2 TSE and FFE images in both quantitative and qualitative analysis. All raters concordantly displayed satisfactory estimation results when assessing spinal canal stenosis using the CycleGAN-generated images with T2 TSE images (kappa = 0.61-0.75) compared to the original FFE with T2 TSE images (kappa = 0.48-0.71).

CONCLUSIONS

CycleGAN demonstrated the capability to produce images with diminished CSF flow artifacts. When paired with T2 TSE images, the CycleGAN-generated images allowed for more consistent assessment of spinal canal stenosis and exhibited agreement levels that were comparable to the combination of T2 TSE and FFE images.

Improving lesion detection in mammograms by leveraging a Cycle-GAN-based lesion remover

BACKGROUND

The wide heterogeneity in the appearance of breast lesions and normal breast structures can confuse computerized detection algorithms. Our purpose was therefore to develop a Lesion Highlighter (LH) that can improve the performance of computer-aided detection algorithms for detecting breast cancer on screening mammograms.

METHODS

We hypothesized that a Cycle-GAN based Lesion Remover (LR) could act as an LH, which can improve the performance of lesion detection algorithms. We used 10,310 screening mammograms from 4,832 women that included 4,942 recalled lesions (BI-RADS 0) and 5,368 normal results (BI-RADS 1). We divided the dataset into Train:Validate:Test folds with the ratios of 0.64:0.16:0.2. We segmented image patches (400 × 400 pixels) from either lesions marked by MQSA radiologists or normal tissue in mammograms. We trained a Cycle-GAN to develop two GANs, where each GAN transferred the style of one image to another. We refer to the GAN transferring the style of a lesion to normal breast tissue as the LR. We then highlighted the lesion by color-fusing the mammogram after applying the LR to its original. Using ResNet18, DenseNet201, EfficientNetV2, and Vision Transformer as backbone architectures, we trained three deep networks for each architecture, one trained on lesion highlighted mammograms (Highlighted), another trained on the original mammograms (Baseline), and Highlighted and Baseline combined (Combined). We conducted ROC analysis for the three versions of each deep network on the test set.

RESULTS

The Combined version of all networks achieved AUCs ranging from 0.963 to 0.974 for identifying the image with a recalled lesion from a normal breast tissue image, which was statistically improved (p-value < 0.001) over their Baseline versions with AUCs that ranged from 0.914 to 0.967.

CONCLUSIONS

Our results showed that a Cycle-GAN based LR is effective for enhancing lesion conspicuity and this can improve the performance of a detection algorithm.