Mouse αSyn aggregates were produced as previously described (Volpicelli-Daley et al., 2014) and kindly provided by Dr. Kelvin Luk (University of Pennsylvania Perelman School of Medicine, United States). Before surgery, αSyn preformed fibrils (PFFs) were sonicated in a water-bath cup-horn sonicator for 4 min (QSonica, Q700 sonicator, 50% power, 120 pulses at 1s ON, 1s OFF) and maintained at room temperature until injection. For control injections, αSyn monomeric protein was maintained on ice until intracerebral injections were performed. Mice were anesthetized with isoflurane and injected unilaterally in the right olfactory bulb (OB) (coordinates from bregma: AP: +5.4 mm; ML: –0.75 mm and DV: –1.0 mm from dura) with 0.8 μL of PFFs or monomeric αSyn (5 μg/μL) for the αSyn induced olfactory dysfunction model and the monomer model respectively; and with matching volumes of vehicle phosphate buffered saline (PBS) as control. Injections were made at a rate of 0.2 μL/min using a glass capillary attached to a 10 μL Hamilton syringe. After injection, the capillary was left in place for five min before being slowly removed. Prior to incision, the animals were injected with local anesthetic into the site of the future wound margin (0.03 mL of 0.5% Ropivacaine; Henry Schein, United States). Following surgery mice received 0.5 mL saline s.c. for hydration, and 0.04 mL Buprenex (Henry Schein) s.c. for analgesia. Mice injected with αSyn into the olfactory system were euthanized 8–15 weeks after surgery.

Stereotactic injections were performed unilaterally into the dorsal striatum (coordinates from bregma: AP: +0.2 mm; ML: ±2.0 mm and DV: –2.6 mm from dura). Experimental animals were injected with 1 μL of mouse PFFs (1 μg/μL) and control animals were injected with 1 μL of vehicle phosphate buffered saline (PBS) at 20–22 weeks of age. Tissue was collected when animals reached 36 weeks of age. Mice were anesthetized by intraperitoneal IP injection with tribromoethanol (Avertin) and perfused transcardially with 0.9% saline, followed by 4% PFA in phosphate buffer. Brains were harvested, post-fixed for 24 h in 4% PFA at 4C, and cryoprotected in 30% sucrose in phosphate buffer for at least 3 days at 4°C or until taken for sectioning. The entire brain of each mouse was cut into 40 μm free-floating coronal sections on a freezing microtome and stored in a cryoprotectant solution consisting of 30% sucrose and 30% ethylene glycol in PBS at –20°C until immunostaining. For detection of the antibody with DAB, we utilized a peroxidase-based method (Vectastain ABC kit and DAB kit; Vector laboratories). Stained sections were mounted onto gelatin-coated glass slides, dehydrated, and coverslipped with Cytoseal 60 mounting medium (Thermo Fisher Scientific).

Z-stacks of mounted Iba1-stained tissue sections were captured at 20x magnification using a whole slide scanner (Zeiss, Axioscan Z1) at a 0.22 μm/pixel resolution. Extended depth focus (EDF) was used to collapse the z-stacks into 2D images as they were collected. Tissue thickness was set to 20 μm, z-stacks were collected at 3 μm intervals and the method setting selected was Contrast. The images were exported with 85% compression as.jpeg files. The digitized images were then uploaded Aiforia^® image processing and management platform (Aiforia Inc, Cambridge, MA, United States) for analysis with deep learning convolutional neural networks (CNNs) and supervised learning.

A supervised, multi-layered CNN was trained using scans of coronal mouse brain tissue sections stained for Iba1, to recognize Iba1+ microglia in mice. The AI model was trained on the most diverse and representative images from across multiple Iba1 datasets and was adapted to each dataset to create multiple AI models capable of accurately detecting Iba1-stained microglia collected by multiple investigators. Whole slide images were loaded to the Aiforia^® cloud for the training data, and each feature layer was trained using tools in the Aiforia^® Cloud platform. All analysis results, including the location of each microglia cell, were determined using the automated deep learning algorithm developed using Aiforia^®. However, the analysis regions (i.e., brain regions; striatum, motor cortex, etc.) were annotated manually.

The first feature layer (“tissue”) was annotated using semantic segmentation to distinguish between Iba1+ tissue and glass slide background. The training regions for the “parent” tissue layer can be described using three tissue training region annotation strategies: positive tissue signal, background, and positive-signal/background interface (edge of tissue). Each training image included at minimum 2 positive tissue signal, 3 background, and 1 positive-signal/background interface training regions. The tissue layer was trained at 10,000 iterations, CNN complexity set to Very Complex, and 100 μm field of view.

Object detection was used for the “child” feature layer (“microglia”) to identify Iba1+ microglia and overall cell diameter. Cell diameters were binned by 10 μm intervals ranging from 10 to 110 μm. Additionally, segmentation was used to distinguish each cell from the surrounding background tissue. On average, 9.5 instance segmentation annotations were included per training image. The training regions for the microglia layer can also be described using four microglia training region annotation strategies: small training regions annotating 2–4 cells, extended background, cell-background, and broad area of interest (Table 1).

Small training regions were primarily used in the early stages of model development, and each image included 1.5 small training regions on average. Cell selection for the small annotations was pseudo counter-balanced across images to include examples of microglia from the olfactory bulb, piriform cortex, anterior olfactory nucleus, dorsal striatum, and substantia nigra. A preliminary model was trained using 1,500 training iterations, and subsequent analyses were performed for 10 of the 35 training images to qualitatively assess model performance. Model performance evaluation was then used to streamline the inclusion of additional annotations, most specifically the broad area of interest training regions.

Broad area of interest training regions were annotated using the Aiforia^® Converter Tool to convert correct AI model predictions that resulted from the previous AI model training round into training annotations for the subsequent round of training. Using this technique, we quickly amplified the number of annotations across a multitude of ROIs that captured staining and scanning variability within the datasets. To maintain a consistent definition for microglia detection, each object converted from analysis results was reviewed and, if necessary, manually adjusted for object diameter and position centering. These annotations made up most of the examples needed for robust microglia identification by supplying 30–800 microglia training examples per training region and, on average, 0.8 broad area of interest training regions were included per training image.

Extended background training region annotations typically included 2–4 cells and extracellular background that extended or “branched” across most of the tissue section in 2–3 directions without including cell bodies or processes. Each branch ranged from 1 to 4 mm, approximately. The aim of the “branching” annotations was to reduce the occurrence of false positives by providing adequate examples of background tissue. On average, 3.6 extended background training regions were included per training image. Finally, to eliminate the occurrence of false positives that result from the incorrect identification of a single microglia as two or more objects, training regions were carefully drawn/edited to exclude any overlapping portions of the object annotations (see Table 1 example images). The final version of this microglia layer completed after 11,789 training iterations, with pre-training instance segmentation network complexity set to ultra complex, and post-training object detection gain and Level of detail adjusted to optimize model performance (object detection gain 2.1, instance segmentation gain 4, and level of detail “High”) (Supplementary Figure 1A).

Previous studies have demonstrated the area/perimeter ratio of microglia as a reliable means for identifying microglia phenotype and the quantification of the area/perimeter ratio using semi-automated methods (George et al., 2019, 2021). Hence, the area/perimeter ratio was included for the comparison of morphology assessment methods, MATLAB and Aiforia^®. First, tissue sections were selected from 4 mice from the αSyn induced olfactory dysfunction mouse model in each of the following brain regions: the inner plexiform layer of the OB, anterior piriform cortex, and dorsal striatum. For each brain region, 3 sections of tissue were selected for inclusion in the comparison dataset and were analyzed by the semi-automated MATLAB script as well as the Aiforia^® Iba1-microglia detection model.

Using Aiforia^®, analysis regions were defined for each serial tissue section to include the entirety of each brain region, as opposed to the three fields of view used for MATLAB (Paxinos and Franklin, 1998), and were then analyzed using the combined tissue plus Iba1-microglia AI model (Figure 1A). Analysis using the semi-automated MATLAB script was performed as previously described (Heindl et al., 2018). In brief, microglia were sampled from each brain region in each mouse using three fields of view per serial tissue section with a total of ∼5–8 cells per field of view.

To determine if there are differences in how each method calculates the area/perimeter ratio, a subset of 5 cells from the OB of each mouse was used for a cell-to-cell paired comparison between assessment methods. Finally, to evaluate the relative efficiency of each method, the timestamps related to dataset acquisition were reviewed and quantified in hours and minutes.

To evaluate how well the microglia detection model can be adapted to additional datasets, images and layer-specific annotations were added to the original training dataset to generate additional microglia detection models. From the original αSyn induced olfactory dysfunction (i.e., “parent”) model, 2 additional detection models were adapted: striatal αSyn aggregation and viral infection. To adapt the model to the new datasets, 3–5 analysis regions were annotated per image for 16 images specific to the viral infection model. Compared to the other datasets, the striatal αSyn aggregation dataset had some variability in background staining and thus 57 analysis regions were annotated across 12 images. Following annotations, a preliminary version of each model was trained using 1,500 training iterations and results were visually reviewed to estimate model performance in tissue detection, microglia detection and microglia segmentation. For the viral infection dataset, there were no errors in tissue detection, and thus no tissue layer annotations were needed. However, to ensure accurate detection of the tissue layer for the striatal αSyn aggregation model, each training image included 3.3 positive tissue signal, 3.9 background and 1 positive-signal/background interface training regions, on average.

To adapt the microglia detection layers, analysis regions were reviewed, and undetected microglia were annotated following the “small training region” strategy, annotating 2–4 microglia per region. To help eliminate the occurrence of false positives, cell-background training regions were annotated to include additional examples of background extracellular space. On average across training images for the viral infection model, 17 cell-background and 40 small training regions were annotated along with 46 microglia, 9 of which were annotated for microglia instance segmentation. To control for differences in background staining throughout the striatal αSyn aggregation dataset, only broad area of interest training regions were used across 7 training images. Finally, for the striatal αSyn aggregation model, 2.6 broad area of interest training regions were annotated across a range of brain regions, with 143 annotated microglia, 3 of which were annotated for microglia instance segmentation, on average across training images. Importantly, to eliminate the occurrence of false positives that result from the incorrect identification of a single microglia as two or more objects, training regions were carefully drawn/edited to exclude any overlapping portions of the object annotations (see Table 1 example training region images).

For the final versions of each model, post-training object detection gain settings, as well as instance segmentation gain and level of detail settings, were adjusted to optimize model performance. The final version of the viral infection model completed after 60,000 training iterations using ultra complex instance segmentation network complexity, post-training object detection gain set to 1.4496, and instance segmentation gain set to 4 with High level of detail. The final version of the striatal αSyn aggregation model completed after 20,000 training iterations using ultra complex instance segmentation network complexity, post-training object detection gain set to 1.25, instance segmentation gain set to 3.75 with High level of detail (Supplementary Figure 1B).

To validate the accuracy of our “parent” microglia detection model (αSyn induced olfactory dysfunction) we compared microglia detection by the AI model with manual annotations of microglia in 80 validation areas across images containing 4-10 microglia from 14 animals excluded from the training data. Validation annotations were manually entered by 4–5 researchers experienced in identifying microglia across a range of functional states. These annotations were then compared to the model’s analysis results (AIforia vs. human), as well as the annotations of each individual researcher (human vs. human), to determine if the model’s performance matched that of human researchers. The features quantified for this comparison include non-microglia detected as microglia (false positive), undetected microglia (false negative), space-dependent overlap of analysis results with human annotations (precision), space-independent overlap of analysis results with annotations (sensitivity), and F-measure (the harmonic mean of precision and sensitivity) (Penttinen et al., 2018; Mäkelä et al., 2021). The AI model was considered valid when the AI vs human results performed equal to or better than human vs human (interobservability) results on average. Next, to demonstrate the limitations of generalizability and the importance of training data in adapting the microglia detection model, each adapted model was validated against manual annotations from researchers, as in the method described above. For validating across independent Iba1+ datasets, the number of validation regions was increased to 330 across 25 images equally representing each of the datasets, as well as 2 additional datasets that were not part of the training data for any of the models.

Data outputs from Aiforia^® were post-processed in Rstudio (ver.1.4.1106) primarily to calculate the area/perimeter ratios as well as the mean object diameter for each analysis region. However, due to background staining variability in the striatal αSyn aggregation dataset, regions with fewer than 100 cells were filtered from the analysis (a total of 4 ROIs removed). Additionally, post-processing was performed to calculate the means of all morphology measures within a single animal for each brain region. Finally, Rstudio Notebook/markdown versions of these scripts are also available to help streamline processing of future analyses (see supporting documents).