Fig 1.
FLLIT system setup and overview of computational workflow.
A. Camera and arena setup used for video capture. B. Segmentation and tracking procedure. i) Training samples are automatically generated, without any user input, by identifying high-confidence leg px (shown in red; located at the intersection between skeletonization and edge morphological operations) and high-confidence non-leg pixels (shown in blue) (see text for details). ii) Training sets are learned and grown by iterative supervised segmentation to derive a classifier. iii) Segmentation of novel images is carried out using the trained classifier. iv) Tracking occurs by matching leg claw positions across adjacent frames. v) Results are given as positions of leg claws in each frame. FLLIT, Feature Learning-based Limb segmentation and Tracking; px, pixels.
Table 1.
Movement and gait data automatically computed by FLLIT include raw body and leg claw position data, as well as 20 leg movement parameters, 5 plots, and a tracked video.
Fig 2.
Accuracy of FLLIT segmentation and tracking results.
A. Representative images of wild-type Drosophila legs taken using the default settings and the manual leg-tip positions identified by two different human users. Blue and green insets are 10 pixels wide and show the respective boxed regions in the top image. Red and yellow dots represent the pixels identified as tip pixels by the two users, within the respective blue and green boxes. B. Frequency distribution of the deviation (in pixels) between leg-tip positions annotated by the two users (n = 54 frames, 324 leg tips, from two videos). Discrepancies can occur in both the x and y directions and are represented as the Euclidean distance between the two pixels. C. Number of corrections required for misidentified legs, normalised to per 1,000 frames (mean = 1.7 corrections; n = 29 videos, 15,166 frames). Plotted as a frequency distribution and a scatter plot (inset). D. Percentage of missing data in wild-type Drosophila after tracking (n = 29 videos, 15,166 frames). E. Frequency distribution of the deviation (in pixels) between computationally and manually derived leg tip positions (n = 106 frames, 636 leg tips from two videos). F. Segmentation F0.5 (P < 0.01), precision (P < 0.05), and recall (P < 0.01) scores improved for each video after learning and application of a FLLIT leg classifier, compared to using morphological parameters alone. (n = 8 videos, 2–3 images per video). P values were calculated using a nonparametric Wilcoxon matched-pairs signed rank test (learned versus morphological for each data set). Bars represent the means and standard deviations. See also S1 Video. Underlying data can be found in S1 Data. FLLIT, Feature Learning-based Limb segmentation and Tracking.
Fig 3.
Gait signatures of Drosophila models of neurodegeneration reveal properties of underlying neuronal dysfunctions.
A. Climbing performance (highest height climbed in 30 s) of flies analyzed. Representative FLLIT-derived walking leg traces of the respective genotypes. C. Cliff’s delta indices of effect sizes (filled circles) of SCA3 and PD-relevant gait parameters from Table 2, with 95% confidence intervals (horizontal lines), with respective P values. Positive Cliff’s delta for a given parameter indicates an increase in mutant flies compared to respective controls, whilst negative Cliff’s delta indicates a decrease. Detailed statistics are given in S1 Table. Raw values are plotted in S5 Fig. The following gait parameters were analyzed: body veering (number of body turns normalized to the average number of strides per leg), footprint regularity (standard deviations of the anterior extreme position, normalized to body length), leg domain lengths (normalized to body length), Average ratio of the hind vs mid domain length of the right and left sides, Domain overlap (number of pixels overlapping between leg domains, normalized to the average number of strides per leg), Stride lengths of the mid and hind legs (normalized to body length), Average ratio of the hind vs mid stride lengths of the right and left sides. Body length measurements were individually obtained for each fly for normalization. Genotypes examined: Elav-Gal4>UAS-SCA3-flQ27 (n = 10), Elav-Gal4>UAS-SCA3-flQ84 (n = 10), Elav-Gal4>+ (n = 9), Elav-Gal4>UAS-SNCA (n = 9), yw (n = 11), park1 (n = 10), mir-263aKO (n = 11), ple-Gal4> UAS-SCA3-flQ27 (n = 14), ple-Gal4> UAS-SCA3-flQ84 (n = 15). D. Climbing performance (highest height climbed in 30 s) of Elav-Gal4>UAS-SCA3-flQ84, Elav-Gal4>UAS-SNCA, and respective UAS/+ control flies. Hashed boxes demarcate the approximately 10 poorest (red boxes) and approximately 10 best (blue boxes) climbers showed in panel E. Genotypes examined: UAS-SCA3-flQ84/+ (n = 10), Elav-Gal4>UAS-SCA3-flQ84 (n = 9 poor climbers, and n = 11 good climbers, the best climbers in the population), UAS-SNCA/+ (n = 11), Elav-Gal4>SCA3-flQ84 (n = 11 poor climbers and n = 12 good climbers). E. Cliff’s delta indices of flies from panel D, comparing the poor and good climbers for Elav-Gal4>UAS-SCA3-flQ84 and Elav-Gal4>UAS-SNCA to their respective UAS/+ controls. P values were calculated using a non-parametric Mann-Whitney test except for park1 and mir-263aKO, which shared the same control (yw); hence, P was calculated using a nonparametric Kruskal–Wallis test with Dunn’s multiple comparisons posthoc test (see S5 Fig). See also Table 2 and S3–S7 Videos. Underlying data can be found in S1 and S2 Data. FLLIT, Feature Learning-based Limb segmentation and Tracking; PD, Parkinsons Disease; SCA3, Spinocerebellar ataxia Type 3; SNCA, alpha-synuclein.
Table 2.
Gait features of PD and SCA3 and corresponding gait parameters computed by FLLIT.
Fig 4.
Detection and characterization of high-frequency leg tremors in Drosophila mutants.
A. Representative leg traces of freely walking control (yw) and Hk2 mutant Drosophila. B. Schematic showing the parameters used to determine tremor events. C, D. Number of shaking (C) and tremor (D) events in control (n = 11), Hk2 (n = 17), and Sh5 (n = 21) Drosophila. E. Distribution of the time interval durations between tremor peaks or valleys in Hk2 flies. A significant proportion of events showed an interval duration of 20–30 ms (P < 0.01; P value was determined by running a nonparametric permutation test with 100,000 iterations), reflecting a tremor frequency of approximately 33–50Hz. F. Top: number of tremors/s in the fore, mid, and hind legs of each Hk2 fly (n = 17). Bottom: percentage of all tremors accounted for by either the fore, mid or hind legs in each Hk2 fly that exhibited tremors. G. Number of tremors per second exhibited by each of the genotypes examined: Elav-Gal4>SCA3-flQ27 (n = 10), Elav-Gal4>SCA3-flQ84 (n = 10), Elav-Gal4>+ (n = 9), Elav-Gal4>SCNA (n = 9), yw (n = 11), park1 (n = 10), mir-263aKO (n = 11), ple-Gal4>SCA3-flQ27 (n = 14), and ple-Gal4>SCA3-flQ84 (n = 15). H. Distribution of the time interval durations between tremor peaks or valleys in Elav-Gal4>SCA3-flQ84 flies. A significant proportion of events showed an interval duration of 20–30 ms (P < 0.0001), reflecting a tremor frequency of approximately 33–50Hz. I. Top: number of tremors/s in the fore, mid, and hind legs of each Elav-Gal4>SCA3-flQ84 fly when (i) walking upright (n = 10) or (ii) walking inverted (n = 15). Bottom: percentage of all tremors accounted for by either the fore, mid, or hind legs in each Elav-Gal4>SCA3-flQ84 fly that exhibited tremors when (i) walking upright (n = 8) or (ii) walking inverted (n = 7). *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001. All data were analyzed using a nonparametric Kruskal–Wallis test with Dunn’s multiple comparisons posthoc test unless otherwise stated above. Bars represent the means and standard deviations. See also S8 and S9 Videos. Underlying data can be found in S1 Data. SCA3, Spinocerebellar ataxia Type 3; SNCA, alpha-synuclein.
Table 3.
Parameters used in the Kernel–Boost algorithm.