aMAP is a validated pipeline for registration and segmentation of high-resolution mouse brain data

Abstract

The validation of automated image registration and segmentation is crucial for accurate and reliable mapping of brain connectivity and function in three-dimensional (3D) data sets. While validation standards are necessarily high and routinely met in the clinical arena, they have to date been lacking for high-resolution microscopy data sets obtained from the rodent brain. Here we present a tool for optimized automated mouse atlas propagation (aMAP) based on clinical registration software (NiftyReg) for anatomical segmentation of high-resolution 3D fluorescence images of the adult mouse brain. We empirically evaluate aMAP as a method for registration and subsequent segmentation by validating it against the performance of expert human raters. This study therefore establishes a benchmark standard for mapping the molecular function and cellular connectivity of the rodent brain.

Figure 1. Anatomical structures used to assess segmentation performance.

(a) An illustration showing the 3D shape of the nine brain structures in the left hemisphere used to assess segmentation performance. Red lines within each structure highlight the coronal plane in the reference atlas that was presented to human raters. (b) For each structure, the segmentation outlines are shown for a given group of 11 raters (grey lines). The consensus outline for the same structure and 11 raters as determined by STAPLE is overlaid (bold coloured line). According to The Allen Brain Atlas nomenclature, the nine structures shown are: anterior cingulate area (ACA); anterior hypothalamic nucleus (AHN); medial vestibular nucleus (MV); retrosplenial cortex (RSP); primary somatosensory area (SSp); subiculum (SUB); primary visual cortex (VISp), secondary visual cortex, anteriomedial part (VISam); ventral posteromedial nucleus of the thalamus (VPM) (D: dorsal; V: ventral; M: medial; L: lateral).

Figure 2. Segmentation performance of human raters and aMAP.

(a) Segmentation outlines of human raters (grey) with the aMAP segmentation result of the same structure and brain overlaid (orange). (b) Dice scores for manual (n=22 raters, each segmenting two of four potential brains, grey) versus aMAP (n=4 brains, orange) segmentations grouped by target structures (n=9). (c) Box plots showing Dice scores of human (grey) versus aMAP (orange) segmentations grouped by brain. Brains used in the registration parameter search (training data) are marked with an asterisk. (d) Cumulative histogram of the Dice scores for manual (grey) and aMAP (orange) segmentations for all structures and brains as shown in b. Vertical lines indicate the median scores.

Figure 3. Sources of variance in manual segmentation.

(a) A box plot showing the anterior–posterior distance between any two human raters (n=22 raters, each segmenting three of six potential brains) in their estimation of the correct optical section (z-choice) for manual segmentation for each brain structure. Structures: anterior cingulate area (ACA); anterior hypthalamic nucleus (AHN); dentate gyrus, granule cell layer (DG-sg); medial vestibular nucleus (MV); retrosplenial cortex (RSP); primary somatosensory area (SSp); subiculum (SUB); primary visual cortex (VISp); secondary visual cortex, anteriomedial part (VISam); ventral posteromedial nucleus of the thalamus (VPM). (b) Example manual segmentations (n=22) of the DG-sg performed by 11 human raters taken from a single test brain and its repeated presentation. (c) Example manual segmentations of the DG-sg, taken from two z-sections from within the data set shown in b. These two z-sections were chosen based on their having multiple segmentation attempts (n=4 outlines shown in each image, left image: anterior, right image: posterior). (d) Schematic highlighting two extreme segmentation reliability scenarios. Bottom left: a given rater may perform poorly against the STAPLE consensus (black square) of all raters (Dice score=0.25, grey lines). However, generation of a Dice score that determines the overlap between the first attempt and the second attempt (intra-rater Dice score, red) indicates high reliability (for example, Dice score=0.98). In contrast, a given rater may obtain a Dice score more similar to the STAPLE consensus but be unreliable in their estimate of the location of the structure (for example, intra-rater Dice score=0.02; top right). (e) Plot of the inter-rater (n=44 segmentations per structure; that is, first and second attempt versus STAPLE consensus for 22 raters per structure, black) and intra-rater (n=22 segmentations, that is, second attempt versus first attempt for 22 raters per structure) Dice scores for each target structure. (f) Plot showing the cumulative histogram of intra- and inter-rater Dice scores for all data presented in e.