Neuron Segmentation using Incomplete and Noisy Labels via Adaptive Learning with Structure Priors

Abstract

Recent advances in machine learning have demonstrated significant success in biomedical image segmentation. Most existing high-quality segmentation algorithms rely on supervised learning with full training labels. However, segmentation is more susceptible to label quality; notably, generating accurate labels in biomedical data is a labor-and time-intensive task. Especially, structure neuronal images are hard to obtain full annotation because of the entangled shape of each structure. In this thesis, a neuron structure semantic segmentation algorithm is proposed on a noise label. I assume that the label has noise and propose two new novel loss functions.Adaptive loss is applied to noise pixels in different labels with prediction in partially annotated labels. These fluorescence images may have confidence that can leverage prior knowledge when each pixel has intensity. Reconstruction loss is suggested that can be regularized of neuronal cell structures to reduce false segmentation near noisy labels. Additionally, This study is aimed to verify that our method preserves the connectivity of linear structure through a novel evaluation matrix.

Deep Learning in Cell Image Analysis

“…An unsupervised learning method for nuclear segmentation in brain images was proposed in [139] to iteratively train a mask R-CNN model with automatically generated noisy instance segmentation masks and refine the labels using an expectation and maximization (EM) procedure. Park et al [140] proposed a robust neuron segmentation method that leveraged ADMSE loss to adaptively reduce the weights of noisy labels. Annotating data by multiple experts improves the quality of labels; however, inconsistency among experts could be a type of noisy label in training models.…”