Abstract
Mitochondria instance segmentation from electron microscopy (EM) images has seen notable progress since the introduction of deep learning methods. In this paper, we propose two advanced deep networks, named Res-UNet-R and Res-UNet-H, for 3D mitochondria instance segmentation from Rat and Human samples. Specifically, we design a simple yet effective anisotropic convolution block and deploy a multi-scale training strategy, which together boost the segmentation performance. Moreover, we enhance the generalizability of the trained models on the test set by adding a denoising operation as pre-processing. In the Large-scale 3D Mitochondria Instance Segmentation Challenge at ISBI 2021, our method ranks the 1st place.
Introduction
Mitochondria are essential organelles, often referred to as the "powerhouses" of the cell. Mapping their 3D structure is critical for understanding cellular function and pathology. However, electron microscopy (EM) produces massive datasets where mitochondria appear as complex, elongated, and densely packed structures. Manual segmentation of these volumes is prohibitively time-consuming.
Automated segmentation faces two primary challenges: anisotropy and noise complexity. EM data often exhibits non-uniform resolution across different axes, and human tissue samples typically contain significantly higher noise levels compared to animal models. To address these, we developed specialized deep residual networks, Res-UNet-R and Res-UNet-H, which secured 1st place in the ISBI 2021 MitoEM Challenge.
Methodology
Our approach utilizes a bottom-up segmentation framework with two distinct architectures tailored to the sample species:
- Res-UNet-R (Rat): A residual U-Net that predicts a semantic mask and instance boundaries through a single decoder path.
- Res-UNet-H (Human): Designed for noisier human data, this variant employs dual decoder paths—one for the semantic mask and another for instance boundaries—to improve feature separation.
Architectural Innovations
A core component of our networks is the Anisotropic Convolution Block (ACB). Standard 3D convolutions often struggle with the non-uniform resolution of EM datasets. The ACB addresses this by using a $1 \times 3 \times 3$ layer followed by cascaded $3 \times 3 \times 3$ layers, specifically targeting the dataset's anisotropic nature.
To further improve robustness, we implemented Multi-scale Training (MT). Models are initially trained on smaller input sizes ($32 \times 256 \times 256$) and subsequently fine-tuned on larger volumes ($36 \times 320 \times 320$). This strategy expands the receptive field and mitigates overfitting. For the noisy human samples, an Interpolation network acts as a pre-processing step, restoring noisy regions using information from adjacent frames.
Results and Discussion
The proposed methods were evaluated on the MitoEM dataset (Rat and Human volumes) and the Lucchi dataset.
Quantitative Performance
Our models achieved state-of-the-art results, outperforming previous benchmarks. On the MitoEM validation set, we achieved AP-75 scores of 0.917 for Rat samples and 0.828 for Human samples. On the Lucchi dataset, the model demonstrated high versatility with a Jaccard index of 0.895 and a Dice score of 0.945.
Qualitative Analysis
Visual inspection confirms the model's ability to accurately delineate individual mitochondria even in densely packed regions. By predicting boundaries alongside semantic masks, the model can effectively generate "seeds" to separate touching instances.
Key Insights
The success of the dual-path decoder in Res-UNet-H suggests that increased noise in human samples requires structural architectural adaptations rather than just additional training data. Furthermore, the denoising pre-processing step proved vital for maintaining high performance in low signal-to-noise environments.
Conclusion
We presented Res-UNet-R and Res-UNet-H, two specialized architectures for 3D mitochondria instance segmentation. By combining anisotropic convolution blocks, multi-scale training, and targeted denoising, our framework achieved top performance in the ISBI 2021 MitoEM Challenge. Future work will focus on integrating end-to-end denoising and automating the tuning of kernel sizes for varying EM resolutions.
Summary updated