Automated detection of potential artifacts in machine learning based bio-image segmentation

Image segmentation algorithms, while powerful, are inherently prone to artifacts, making perfect segmentation theoretically and practically impossible. We propose an automated artifact identification scheme for posterior rapid manual re-correction to address this challenge. Hereby, our contribution is twofold: We extend our previous work, delivering polynomial defenses (PDs). These defenses mimic noise distributions that significantly improve segmentation quality when removed from the training images. In practice, we perturb unseen images with PDs and demonstrate that the resulting segmentation differences achieve promising precision in artifact detection compared to traditional Gaussian and Poisson noise perturbations. This automated guidance is our essential contribution. Beyond improving the reliability of image-processing outputs, our approach provides a valuable tool for enhancing manually segmented training datasets. Hereby, the automated guidance massively decreases manual cross-checking time.