Multitask learning via pseudo-label generation and ensemble prediction for parasitic egg cell detection: IEEE ICIP Challenge 2022
Issued Date
2022-01-01
Resource Type
ISSN
15224880
Scopus ID
2-s2.0-85136124623
Journal Title
Proceedings - International Conference on Image Processing, ICIP
Start Page
4273
End Page
4277
Rights Holder(s)
SCOPUS
Bibliographic Citation
Proceedings - International Conference on Image Processing, ICIP (2022) , 4273-4277
Suggested Citation
Aung Z.H., Srithaworn K., Achakulvisut T. Multitask learning via pseudo-label generation and ensemble prediction for parasitic egg cell detection: IEEE ICIP Challenge 2022. Proceedings - International Conference on Image Processing, ICIP (2022) , 4273-4277. 4277. doi:10.1109/ICIP46576.2022.9897464 Retrieved from: https://repository.li.mahidol.ac.th/handle/20.500.14594/84378
Title
Multitask learning via pseudo-label generation and ensemble prediction for parasitic egg cell detection: IEEE ICIP Challenge 2022
Author(s)
Author's Affiliation
Other Contributor(s)
Abstract
Parasitic infections are one of the leading causes of deaths and other ailments worldwide. Detecting such infections using traditional diagnostic procedures requires experienced medical technologists together with a significant amount of time and effort. An automated procedure with the ability to accurately detect parasitic diseases can greatly accelerate the process. This work proposes a deep learning-based object detection for parasitic egg detection and classification. We show that multitask learning via pseudo-mask generation improves the single model performance. Moreover, we show that a combination of multitask learning, pseudo-label generation, and ensembling model predictions can accurately detect parasitic egg cells. Continuous training via pseudo-label generation and ensemble predictions improves the accuracy of single-model detection. Our final model achieved a mean precision score (mAP) of 0.956 on a validation set of 1, 650 images. Our best model obtained mIoU and mF1 scores of 0.934 and 0.988 respectively. We discuss its technical implementation in this paper.