Real-time Automatic M-mode Echocardiography Measurement with Panel Attention from Local-to-Global Pixels

This is the implemented code of RAMEM, Real-time Automatic M-mode Echocardiography Measurement with Panel Attention from Local-to-Global Pixels. This project is modified based on Yolact_minimal. This project covers:

• MEIS, a dataset of M-mode echocardiograms for instance segmentation.
• Panel attention, a depth-wise local-to-global efficient attention by pixel-unshuffling, embedding with updated UPANet V2, updated upon UPANet V1.
• AMEM, an efficient algorithm for targeting fast and accurate automatic labelling among diagnoses.

(AV) (LV) 1

Preparation

Please follow Yolact_minimal to set up the needed environment.

Datasets

Opened datasets - PASCAL & COCO

Please follow the same instruction at Yolact_minimal.

Proposed dataset - MEIS

Please find MEIS (the M-mode echocardiograms dataset in COCO format) at link. The main images are in ./raw_image ². The annotations are Allmode_train.json and Allmode_test.json. There also are human labelling results in ./data/manual_results.xlsx and ground truth raw information in ./data/df_dict. The distribution of the dataset is below.

Panel attention & UPANets V2

Another core of this work is proposing a new novel efficient non-local attention in UPANets as UPANet V2. The panel attention can be fetched in ./modules/upanets_v2.py

AMEM

Finally, the last core of this work is an automatic M-mode echocardiography measurement, AMEM. The algorithm can be seen below with implemented code in ./detect.py. For different types in av and lv, functions of measure_av and measure_lv in ./detect.py are designed accordingly.

Weights

There are six weight groups in this project from UPANet V2 (the proposed backbone) in maYOLACT to existing works or different backbones (works: maYOLACT and RTMDet, backbones: maDarkNet and SwimTransformer_Tiny). Please fetch each weight according to the number link #.

Implementation

The primary implementation is ./main.py, which allows to switch between training, evaluation, and detection. Please note that this project only supports one GPU operation.

Training

End-to-end

# Basic training UPANet80 V2 with default setting from end-to-end on PASCAL 2012 SBD,
python main.py --env_path='/ur_env_path' --train=True --resume=False --cfg=upanet80_pascal -- train_bs=8

# Basic training UPANet80 V2 with default setting from end-to-end on MEIS, train_bs=9 is to prevent error in random batch size with no ground truth bounding box after data augmentation in the last one data.
python main.py --env_path='/ur_env_path' --train=True --resume=False --cfg=upanet80_meis -- train_bs=9

Resume (Fine-tuning)

python main.py --env_path='/ur_env_path' --train=True --resume=True --cfg=upanet80_* -- train_bs=* --weight='/ur_weight_path'

Evaluation

Evaluation w/o measurement (in MEIS)

If --train=True and --plot_every_iters=1, the evaluation will be implemented every epoch.

python main.py --env_path='/ur_env_path' --train=True --plot_every_iters=1 --resume=True --cfg=upanet80_* -- train_bs=*

If a direct evaluation is needed, set --train=False and resume=True along with --weight='/ur_weight_path' will do.

python main.py --env_path='/ur_env_path' --train=False --resume=True --cfg=upanet80_* --weight='/ur_weight_path'

Evaluation w/ measurement (in MEIS)

If AMEM is wanted upon M-mode echocardiography evaluation, please set --measure==True, --scaler=True for auto scaler detecting from OCR, --scaler=* for manual setting scaler, --eval_humans=True for calculating error compared with humans results.

python main.py --env_path='/ur_env_path' --train=False --resume=True --cfg=upanet80_* --weight='/ur_weight_path' --measure==True --scaler=True --eval_humans=True 

Detection

Image detection

This project can automatically label a pile of images from a folder in --image for general labelling upon captured images from ultrasound machines.

python main.py --env_path='/ur_env_path' --train=False --resume=True --cfg=upanet80_* --weight='/ur_weight_path' --measure==True --scaler=True --eval_humans=False --image='/ur_image_folder_path' --image_detect=True

The AV (left) and LV (right) labelling results based on prediction mask in a real-time instance segmentation can be as the following.

Video detection

There are two routes to get the detected results for real-time detection via video. One is directly showing the result through cv2 windows, which could slow down the speed to slightly under real-time. By setting --video, --video_detect=True, and --show_img=True

python main.py --env_path='/ur_env_path' --train=False --resume=True --cfg=upanet80_* --weight='/ur_weight_path' --measure==True --scaler=True --eval_humans=False --video='/ur_video_folder_path' --video_detect=True --show_img=True

Another one is to scan video in the background without showing through cv2 directly, which will need to set --save_image=True to save each detected frame result as a temporary image. And then, the frames will be collected into a video. The demo above shows a real-time detection through input to the detection result ³.

python main.py --env_path='/ur_env_path' --train=False --resume=True --cfg=upanet80_* --weight='/ur_weight_path' --measure==True --scaler=True --eval_humans=False --video='/ur_video_folder_path' --video_detect=True --save_img=True

Note

¹ The demo is generated by AMEM (our proposed algorithm) upon the video.

² In /raw_image, the file name follows a policy that the first number name is the raw image and the second number name is ground truth with medical manual labelling (instead of .json labelling).

³ For declaration, by the example of YOLACT, the real-time calculation is from input to NMS (Non-maximum suppression). The general agreement is either >30 or >24 FPS. However, this coverage is certainly not practical in real-world detection, as it could ignore the full-stack operation computational overhead. Here, we have no intention to cover or fix this dilemma as well. This work, instead, only shows the possibility of making a real-time (taking >24 FPS) detection from input to detection results (after M-mode echocardiography measurement) and nearly real-time capability in showing the detection directly. An improvement in speed and optimisation is definitely needed in the future.