Brief Review — Parcellation of Visual Cortex on High-Resolution Histological Brain Sections Using Convolutional Neural Networks

Atlas-Aware Net, Better than Base Net (Which Based on U-Net)

Parcellation of Visual Cortex on High-Resolution Histological Brain Sections Using Convolutional Neural Networks,
Atlas-Aware Net, by Forschungszentrum Jülich, and Heinrich Heine University Düsseldorf
2017 ISBI, Over 20 Citations (Sik-Ho Tsang @ Medium)
Medical Image Analysis, Biomedical Image Segmentation

• An automatic approach is proposed for parcellating histological sections at 2μm resolution, via CNN which combines topological information from probabilistic atlases with the texture features learned from high-resolution cell-body stained images.

Outline

1. Atlas-Aware Net
2. Results

1. Atlas-Aware Net

1.1. Base Net (Baseline)

Base Net Bsaed on U-Net

• Base net, based on U-Net, is constructed, consisting of 10 blocks containing several layers (denoted by colored rectangles). The number of channels of the layers is shown at the bottom of each block. Batch normalization and ReLU are applied after each convolution.
• Total number of parameters: 1479728. Receptive field size: 1481² px² ≈ 3²mm². The output resolution of our network to 16μm.

1.2. Atlas-Aware Net (Proposed)

Atlas-Aware Net

• To identify a certain area in the brain, neuroscientists first identify a region of interest (ROI) on a low-resolution global view and then map the precise borders of this area by means of local texture patterns on a high-resolution local view.
• In a similar manner, the proposed model combines precise local features from high-resolution cell-stained tissue with a relatively imprecise but topologically correct probabilistic atlas prior.

Atlas-aware net includes transformed probabilistic atlas (atlas prior) in an additional contracting path. Activations from both contracting paths are appended to the joint expansive path (dashed lines). Input block 1' for the atlas prior consists out of one 5×5×16 convolution (no pooling).

• The network learns the atlas prior faster than the image of the cell-body stained section, because the atlas data are less complex and directly represent an estimate of the output labels.

2. Results

Gm/wm/bg segmentation, with input image left and segmentation result right. Solid lines represent manual delineations of inner (black) and outer (white) cortical boundary

• Compared to expert segmentations, the automatic segmentation is consistent and reproducible, but seems to systematically underestimate the inner cortical boundary.

Quantitative evaluation of base and atlas-aware model. Label numbers 14–16 denote labels gm, wm, and bg.

• (b): The base model predicts the most frequent areas (hOc1, hOc2, hOc4la) with good precision.
• (c): Yet, only the atlas-aware model manages to predict areas which are not as much represented in the dataset.

Qualitative evaluation of base and atlas-aware models.

• (a): annotated areas (color) and projected probability of area hOc1 from the atlas (gray),
• (b): segmentation with the base model (no atlas prior), and
• (c)-(e): segmentations with the atlas-aware model on consecutive sections.

Reference

[2017 ISBI] [Atlas-Aware Net]
Parcellation of Visual Cortex on High-Resolution Histological Brain Sections Using Convolutional Neural Networks

1.10. Biomedical Image Segmentation

2015 … 2017 [Atlas-Aware Net] … 2021 [Ciga JMEDIA’21]

My Other Previous Paper Readings