DDeep3M+ : adaptive enhancement powered weakly supervised learning for neuron segmentation

Abstract

Significance: Robust segmentations of neurons greatly improve neuronal population reconstruction, which could support further study of neuron morphology for brain research. Aim: Precise segmentation of 3D neuron structures from optical microscopy (OM) images is crucial to probe neural circuits and brain functions. However, the high noise and low contrast of images make neuron segmentation challenging. Convolutional neural networks (CNNs) can provide feasible solutions for the task but they require large manual labels for training. Labor-intensive labeling is highly expensive and heavily limits the algorithm generalization. Approach: We devise a weakly supervised learning framework Docker-based deep network plus (DDeep3M+) for neuron segmentation without any manual labeling. A Hessian analysis based adaptive enhancement filter is employed to generate pseudo-labels for segmenting neuron images. The automated segmentation labels are input for training a DDeep3M to extract neuronal features. We mine more undetected weak neurites from the probability map based on neuronal structures, thereby modifying the pseudo-labels. We iteratively refine the pseudo-labels and retrain the DDeep3M model with the pseudo-labels to obtain a final segmentation result. Results: The proposed method achieves promising results with the F1 score of 0.973, which is close to that of the CNN model with manual labels and superior to several segmentation algorithms. Conclusions: We propose an accurate weakly supervised neuron segmentation method. The high precision results achieved on 3D OM datasets demonstrate the superior generalization of our DDeep3M+.