• Lung Nodules Dectection
• Description
• Model Architecture
• Dataset
• Environment Requirements
• Quick Start
• • • Result
• Model Description
• • Performance
  • • Training Performance
    • Inference Performance On Ascend 910
    • Inference Performance On Ascend 310

Lung Nodules Dectection

Description

We use CenterNet as backbone to detect lung nodules. CenterNet is a novel practical anchor-free method for object detection, 3D detection, and pose estimation, which detect identifies objects as axis-aligned boxes in an image. The detector uses keypoint estimation to find center points and regresses to all other object properties, such as size, 3D location, orientation, and even pose. In nature, it's a one-stage method to simultaneously predict center location and bboxes with real-time speed and higher accuracy than corresponding bounding box based detectors.

Paper Setio AA, Traverso A, De Bel T, Berens MS, Van Den Bogaard C, Cerello P, Chen H, Dou Q, Fantacci ME, Geurts B, van der Gugten R. Validation, comparison, and combination of algorithms for automatic detection of pulmonary nodules in computed tomography images: the LUNA16 challenge. Medical image analysis. 2017 Dec 1;42:1-3.

Model Architecture

The stacked Hourglass Network downsamples the input by 4×,followed by two sequential hourglass modules.Each hourglass module is a symmetric 5-layer down-and up-convolutional network with skip connections .This network is quite large ,but generally yields the best keypoint estimation performance.

Dataset

Note that you can run the scripts based on the dataset mentioned in original paper or widely used in relevant domain/network architecture. In the following sections, we will introduce how to run the scripts using the related dataset below.

Dataset used: LUNA16 It consists of 1,186 lung nodules annotated in 888 CT scans.

• Dataset size：26G

  • Train：19G，118000 images
  • Val：0.8G，5000 images
  • Test: 6.3G, 40000 images
  • Annotations：808M，instances，captions etc

• Data format：image and json files

• Note：Data will be processed in dataset.py

Environment Requirements

• Hardware（Ascend）

  • Prepare hardware environment with Ascend processor.

• Framework

  • MindSpore

  pip install -r requirements.txt
  

• Dataset should be organize as COCO format and use convert_dataset_to_mindspore.sh to convert data to mindspore.

Quick Start

PATH in shell may need to update according to your dataset location.

```shell
# create dataset in mindrecord format
bash scripts/convert_dataset_to_mindrecord.sh 

# standalone training on Ascend
bash scripts/run_standalone_train_ascend.sh 

# eval on Ascend
bash scripts/run_standalone_eval_ascend.sh 
```

Result

Inference

 #eval_txt.log
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.513
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.938
 Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.507
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.508
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.593
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = -1.000
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.586
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.602
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.602
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.599
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.631
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = -1.000

Model Description

Performance

Training Performance

CenterNet on 11.8K images(The annotation and data format must be the same as coco)

Parameters	CenterNet_Hourglass
Resource	Ascend 910; CPU 2.60GHz, 192cores; Memory, 755G
uploaded Date	3/27/2021 (month/day/year)
MindSpore Version	1.3.0
Dataset	LungDetection
Training Parameters	8p, epoch=130, steps=158730, batch_size = 12, lr=2.4e-4
Optimizer	Adam
Loss Function	Focal Loss, L1 Loss, RegLoss
outputs	detections
Checkpoint	2.1G (.ckpt file)

Inference Performance On Ascend 910

CenterNet on validation(5K images) and test-dev(40K images)

Parameters	CenterNet_Hourglass
Resource	Ascend 910; CPU 2.60GHz, 192cores; Memory, 755G
uploaded Date	11/27/2021 (month/day/year)
MindSpore Version	1.1.0
Dataset	LungDetection
batch_size	1
outputs	mAP
Accuracy(validation)	MAP: 41.5%, AP50: 60.4%, AP75: 44.7%, Medium: 45.7%, Large: 53.6%

Inference Performance On Ascend 310

CenterNet on validation(5K images)

Parameters	CenterNet_Hourglass
Resource	Ascend 310; CentOS 3.10
uploaded Date	11/31/2021 (month/day/year)
MindSpore Version	1.3.0
Dataset	LungDetection
batch_size	1
outputs	mAP
Accuracy(validation)	MAP: 51.3%