Towards Explainable Automated Neuroanatomy

TL;DR

The paper tackles automated neuroanatomy structure identification by shifting from texture-based, black-box CNN approaches to an interpretable, neuron-centric framework built on cell-shape features. It introduces a pipeline combining OpenCV segmentation, diffusion-map-based cell representations, and region-level statistics with 26 binary XGBoost detectors to identify brain structures, aiming for interpretability and cross-modality robustness. The approach achieves strong accuracy (average ROC AUC ~$0.892$ vs $0.924$ for CNN baselines) and produces probability maps that generalize across staining methods without retraining, with explanations grounded in feature importances. This work advances practical connectomic mapping by providing anatomist-friendly, texture-robust, and interpretable structure detection that can align brains across individuals and staining protocols.

Abstract

We present a novel method for quantifying the microscopic structure of brain tissue. It is based on the automated recognition of interpretable features obtained by analyzing the shapes of cells. This contrasts with prevailing methods of brain anatomical analysis in two ways. First, contemporary methods use gray-scale values derived from smoothed version of the anatomical images, which dissipated valuable information from the texture of the images. Second, contemporary analysis uses the output of black-box Convolutional Neural Networks, while our system makes decisions based on interpretable features obtained by analyzing the shapes of individual cells. An important benefit of this open-box approach is that the anatomist can understand and correct the decisions made by the computer. Our proposed system can accurately localize and identify existing brain structures. This can be used to align and coregistar brains and will facilitate connectomic studies for reverse engineering of brain circuitry.

Figures (7)

• Figure 1: The pipelines of our proposed system, where the figures indicate the outputs of the stages and Roman letters enumerate the procedures. I: OpenCV for cell segmentations. II: K-means, Diffusion Mapping, and feature space alignment for extracting cell features. III: Cumulative CDFs for generating regional features. IV: XGBoost for the classification of the region.
• Figure 1: (a) Original cell patches, (b) Cell patches selected using K-means after normalizing rotation.
• Figure 2: (a) Visualization of patch clouds from two brains with different modalities on the 1st and 4th fixed features. Each "dot" in the image is a cell patch. The red patches are from a brain that is stained with thionin and imaged using brightfield, and the blue patches are from a brain that is stained with NeuroTrace blue and imaged using fluorescence. (b) The region of very small cells. (c) The region of large and round cells. (d) The region of thin cells.
• Figure 3: Cell shape distributions for two regions near the Abducens Nucleus (6N). The left image highlights a region that roughly corresponds to the 6N structure (circled in green), while the right image shows a region outside the structure (circled in red). The CDF graph represents the rotation feature of cells, with the green curve depicting the CDF for cells within region 1 and the red curve for those within region 2.
• Figure 4: Comparison of ROC AUC scores for detecting different structures. Please see chen2019active for the list of all 26 structures and their abbreviations.