Supervised Versus Self-Supervised: Which is Better for Biomedical Image Segmentation?

Brouwer, Eric (2022) Supervised Versus Self-Supervised: Which is Better for Biomedical Image Segmentation? Bachelor's Thesis, Artificial Intelligence.

Preview	Text bAI_2022_Eric_Brouwer.pdf Download (12MB) | Preview
	Text toestemming.pdf Restricted to Registered users only Download (125kB)

Abstract

Self-supervised learning (SSL) is a promising key to solving this issue through feature learning under a pretext task, which is then transferred to a downstream main task - in our case image segmentation. Regarding this task, a ResNet50 U-Net was first trained to restore images of liver progenitor organoids from augmented images obtained by random pixel drop, blurring, and sobel filtering. Using a Structural Similarity Index Metric (SSIM) loss as well as the SSIM combined with Mean Absolute Error (L1) loss, both encoder and decoder were trained in tandem. The weights were transferred to another U-Net designed for segmentation with frozen encoder weights, where they were trained with the Binary Cross Entropy, Dice, and Jaccard loss. Paired with this, we also used the same U-Net to train two supervised models, one utilizing the ResNet50 encoder, and the other a simple CNN. Results showed that SSL models using a 25% pixel drop or image blurring augmentation performed better in comparison to the other augmentation techniques paired with the Jaccard loss. When trained on 114 images for the main task, the SSL approach outperforms the supervised method achieving an F1-score of 0.85 with higher stability, in contrast to the 0.78 scored by the supervised method. Furthermore, when trained with larger data sets (1.000 images), SSL is still able to outperform the supervised achieving an F1-score 0.92, contrasting to the score of 0.85 for the supervised method.

Item Type:	Thesis (Bachelor's Thesis)
Supervisor name:	Haja, A. and Schomaker, L.R.B.
Degree programme:	Artificial Intelligence
Thesis type:	Bachelor's Thesis
Language:	English
Date Deposited:	11 Aug 2022 12:42
Last Modified:	11 Aug 2022 12:42
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/28350

Actions (login required)

View Item