Intervertebral Disc Labeling With Learning Shape Information, A Look Once Approach

Labeling vertebral discs from MRI scans is crucial for the proper assessment of spine-related diseases. Challenges such as complex background of MRI images, the similarity between discs and bone area to name a few, usually exacerbates the notorious problem in segmentation of vertebral discs. To overcome this issue, we propose to incorporate shape information within the learning process. Moreover, as labelling anatomical structures such as intervertebral discs usually produces both false positive (FP) and false negative (FN) detections, we propose a look-once approach for the post-processing step in the intervertebral disc labeling procedure.

If this code helps with your research please consider citing the following paper:

R. Azad, Moein Heidari, Ehsan Adeli, Julien Cohen-Adad and Dorit Merhof , "Intervertebral Disc Labeling With Learning Shape Information, A Look Once Approach", download link.

@article{azad2022intervertebral,
  title={Intervertebral Disc Labeling With Learning Shape Information, A Look Once Approach},
  author={Azad, Reza and Heidari, Moein and Cohen-Adad, Julien and Adeli, Ehsan and Merhof, Dorit},
  journal={arXiv preprint arXiv:2204.02943},
  year={2022}
}

Please consider starring us, if you found it useful. Thanks

Updates

• April 7, 2022: First release (Complete implemenation for Spine Generic Dataset added)

Prerequisties and Run

This code has been implemented in python language using Pytorch libarary and tested in ubuntu, though should be compatible with related environment. The required libraries are included in the requiremetns.txt file. Please follow the bellow steps to train and evaluate the model.

1- Download the Spine Generic Public Database (Multi-Subject).

2- Run the create_dataset.py to gather the required data from the Spin Generic dataset.

3- Run prepare_trainset.py to creat the training and validation samples.

Notice: To avoid the above steps we have provided the processed data for all train, validation and test sets here

(should be around 150 MB) you can simply download it and continue with the rest steps. Please unzip the file in the prepared_data folder.

4- Run the main.py to train and evaluate the model. It only takes couple of hours to train with 5GB GPU memory. Use the following command with the related arguments to perform the required action:

A- Train and evaluate the model python src/main.py. You can use --att true to use the attention mechanisim.

B- Evaluate the model python src/main.py --evaluate true it will load the trained model and evalute it on the validation set.

C- You can run make_res_gif.py to creat a prediction video using the prediction images generated by main.py for the validation set.

D- You can change the number of stacked hourglass by --stacks argument. For more details check the arguments section in main.py.

5- Run the test.py to evaluate the model on the test set alongside with the metrics.

Quick Overview

Detailed structure of the proposed shape attention mechanism.

Perceptual visualization of the proposed post-processing approach to eliminate the rate of FP / FN detections.

Results

Our analysis was based on the publicly available Spine Generic Dataset. In bellow, results of the proposed approach illustrated.

Table 1 : Intervertebral disc labeling results on the spine generic public dataset (T1 Modality). Note that DTT indicates Distance to target

Methods	DTT (mm)	FNR (%)	FPR (%)
Ullmann et. all Template Matching	1.97(±4.08)	8.1	2.53
Rouhier et. all Countception	1.03(±2.81)	4.24	0.9
Azad et. all Pose Estimation	1.32(±1.33)	0.32	0.0
Baseline Proposed	1.45(±2.70)	7.3	1.2
Azad et. all Proposed	1.2(±1.90)	0.7	0.0

Table 2 : Intervertebral disc labeling results on the spine generic public dataset (T2 Modality). Note that DTT indicates Distance to target

Methods	DTT (mm)	FNR (%)	FPR (%)
Ullmann et. all Template Matching	2.05(±3.21)	11.1	2.11
Rouhier et. all Countception	1.78(±2.64)	3.88	1.5
Azad et. all Pose Estimation	1.31(±2.79)	1.2	0.6
Baseline Proposed	1.80(±2.80)	5.4	1.8
Azad et. all Proposed	1.28(±2.61)	0.9	0.0

Intervertebral Disc Labeling result on test data

(a): Intervertebral labeling results of three representative T2 images. upper row: ground truth, lower row: predictions. (b): Before (left) and after (right) applying look-once approach on the T1 generated noisy prediction.

Results of the proposed post-processing approach

Table 3 : Performance comparison of the proposed post-processing approach vs the SOTA approach

for eliminating FP detection.

Methods	F1	Accuracy	Specificity	Sensitivity	AUC
Rouhier et. all Condition based	0.850	0.881	0.891	0.902	0.890
Azad et. all Search tree	0.902	0.921	0.925	0.914	0.920
Proposed without geometrical relationship module	0.914	0.932	0.941	0.917	0.9292
Proposed Only look once	0.942	0.958	0.967	0.942	0.955

Comparing inference time of the proposed method vs the search-tree based approach

For more results of the proposed method for intervertebral disc labeling please refer to the paper

Query

All implementations are done by Reza Azad and Moein Heidari. For any query please contact us for more information.

rezazad68@gmail.com
moeinheidari7829@gmail.com