Sartorius---Cell-Instance-Segmentation

https://www.kaggle.com/competitions/sartorius-cell-instance-segmentation

• This is a Code Competition.
• GPU Notebook <= 9 hours run-time.
• Internet access disabled.

NCKU "Machine Learning" - (Competition)

Competition Description

For this semester's final project, our focus is on the instance segmentation competition for neural cells hosted by the internationally renowned biotechnology and pharmaceutical company, Sartorius, on Kaggle [1]. Neurological disorders remain a leading cause of global mortality and disability. The lack of effective quantitative methods for assessing the therapeutic efficacy of diseases poses challenges in devising treatment plans and developing new drugs. Traditional examination methods involve observing the morphology of neural cells through optical microscopes, but the vast number of neural cells makes precise judgment and quantification by human efforts challenging.

Therefore, AI assistance is crucial for analysis, and the initial step in AI analysis involves segmenting individual neural cells. Accurate segmentation of different neural cells enables quantitative data analysis with the aid of AI. Researchers can measure the impact of therapeutic drugs on neural cells, facilitating the advancement of new drug development.

The primary objective of this competition is to address the diverse appearance changes in the most challenging SH-SY5Y type neural cells.

Dataset Download

kaggle competitions download -c sartorius-cell-instance-segmentation

In order to prevent the database from being deleted, I first backed it up to my personal cloud space.

• (Dataset-GoogleDrive Link) & (Annotation-json_file)

Dependencies

• pytorch-21.06-py3:latest

sudo apt-get update
sudo apt-get install ffmpeg libsm6 libxext6  -y
pip install pycocotools
pip install 'git+https://github.com/facebookresearch/detectron2.git'
pip install fastcore
pip install ensemble-boxes
pip install nms

Methods

Detailed description of each part is in PDF File.

1. Mask R-CNN
2. Model Ensembling
3. Image Normalizarion
4. Augmentation
5. Test Time Augmentation

code

showimg.py

• Check the images and the annotation inside the dataset.

python showimg.py

train.py

• training the model.

python train.py

inference.py

• Inference the model with the final model .pth file.

python inference.py

ensemble_inference.py

• Ensemble multiple models or difference test time augmentation.

python ensemble_inference.py

Experiment Results

Detailed description of each table is in PDF File.

1. Sartorius Cell Segmentation Ablation Results

NO	Backbone	Training Iteration	Model Ensemble	Imagenet Normalization	Color Augmentation	Test Time Augmentation	mAP
1	ResNet50	4,000	-	-	-	-	0.285
2	ResNet50 + ResNeXt101	4,000	v	-	-	-	0.291
3	ResNet50 + ResNeXt101	4,000	v	v	-	-	0.293
4	ResNet50 + ResNeXt101	4,000	v	-	v	-	0.288
5	ResNet50 + ResNeXt101	4,000	v	-	-	v	0.25

2. Sartorius Cell Segmentation Results

NO	Backbone	Training Iteration	Model Ensemble	mAP
1	ResNet50	9,679	-	0.300
2	ResNet50	10,679	-	0.302
3	ResNet50 + ResNet50	10,679	v	0.303
4	ResNet50 + ResNet50	11,679	v	0.304

Competition Result

• Our submitted model achieved an mAP of 0.316, securing the 179th position out of 1506 on the competition leaderboard.

• The mAP of the top five can reach about 0.35.

Conclusion and Future work

By employing model ensembling with two Mask R-CNN models trained over an extended period, the final model achieved an mAP of 0.304 for neural cell instance segmentation. In the future, it may be worthwhile to consider incorporating a semi-supervised approach by introducing additional unlabeled images for training. This strategy would enable the model to learn better neural cell image features from a more diverse set of data, potentially leading to increased accuracy.