Entomopathogenic nematodes (EPNs) are organisms that are often mass-produced as biological control agents (BCAs) to mitigate pesticide-related hazards and foster environmental sustainability. Enhancing our understanding of EPNs biology and their interactions with hosts is crucial for refining the use of EPNs in integrated pest management. This study pioneers an interdisciplinary approach, integrating engineering, and entomology, to investigate the behavior of Steinernema carpocapsae EPN. Employing a novel blend of deep learning and optical flow analysis, a Convolutional Neural Network (CNN) efectively recognizes how nematodes react to host-borne stimuli. Achieving remarkable precision (1) and an overall accuracy of 0.938, the model elucidates EPN behaviors in a prototyped microfluidic platform reproducing a host-environment context. The integration of optical flow analysis highlights an increased motor activity in EPNs when exposed to stimuli, adding novel information on their dynamic responses. This versatile methodology represents a significant advancement in detecting and understanding EPNs responses to diverse stimuli, fostering their use as advanced BCAs in sustainable pest control and environmental management.
minimal turbulent motion. A deep learning approach and then fabricated by rapid prototyping in a biocomdiferentiated between scenarios involving stimulus pres- patible resin (VisiJet® M3 Crystal, 3D Systems). The ence and control conditions, enabling the analysis of microfluidic testing arena, featuring a releasing chamber motor diferences attributed to the stimulus presence. (Ø=5mm; h=2mm) and two stimuli chambers (Ø=3mm; Nematodes are utilized as biosensors to gather environ- h=2mm), creates small environments for the analysis of mental information and investigate changes in behavioral nematode behavior without imposing spatial constraints traits in response to stimuli. The integration of optical on their motility, thereby avoiding potential biases. Each lfow analysis revealed distinct variations in motor ac- stimuli chamber is connected to the releasing chamber tivity among individuals across diferent contexts. This through an aisle creating an y-maze arena. The floor of study bridges microfluidics, machine learning, and sus- the test arena is represented by a transparent glass plate tainable integrated pest management, ofering insights ifrmly connected to the base of the upper component by into S. carpocapsae EPN behavior. It contributes to the depositing and curing a polydimethylsiloxane (PDMS) broader context of the convergence of AI, engineering, film (Sylgard 184). To observe and record the behavior of and entomology, advancing strategies for biomedical re- nematodes, an operator introduced the nematodes into search and sustainable pest control. the releasing chamber using a pipette. Subsequently, the miniaturized testing arena was carefully positioned under a 3D visual inspection microscope (Nikon TMS) with 2. Methods a total magnification of 25X. This setup allowed for precise and detailed observation of EPN behavior within 2.1. Steinernema carpocapsae the microfluidic system, ensuring accurate recording of their responses to the specific stimuli provided in the chambers.
We used NEMOPAK SC, a commercial product distributed by BIOPLANET (Cesena, Italy), to obtain and test Steinernema carpocapsae nematodes at infective juvenile stage (IJ). Nematodes were maintained at 4-8°C till the expira- 2.4. Experiment workflow and recordings tion date and solubilized in water at 25°C, with constant stirring and heating, obtained using a magnetic vortex, to keep them viable before trials.
The nematodes were reactivated in water at a temperature of 25°C for 20 minutes as arranged by Bioplanet srl. (Cesena, Italy). Meanwhile, the microarena was washed with water and 70% ethanol to remove contaminants, re2.2. Tested cue duce surface tension, enhance arena wettability, and imEPNs are attracted by chemical cues deriving from po- prove nematode locomotion. A second washing was cartential hosts. Some studies show the attractive efects of ried out using water to eliminate the alcohol traces potengut content of hosts on EPN locomotion [13]. Thus, we tially harmful to the EPN. The chip was then filled with used Lobesia botrana 5th instar larvae feces to investigate the medium, 0.9% NaCl physiological solution within responses in S. carpocapsae. which the nematodes move, it is useful to avoid osmotic stress. Solution and nematodes were inserted using a 2.3. Microfluidic platform and plastic micro pipette. The chip thus assembled is placed under the microscope and the stimulus is positioned in experimental apparatus the opposite arenas, alternating localization to avoid bias during the analysis. The videos were recorded under the microscope for a total of 5 minutes. Video analysis The miniaturized testing arena was designed in SolidWorks (Dassault Systemes, Velizy Villacoublay, France) considers only 60 seconds recording to avoid acclima- sured within each subset. Manual labeling of images was tion time and promote next computer vision and deep performed using the online Makesense software. The learning analysis. Videos are recorded by using an RGB dataset also includes approximately 10% background imcamera (48MP, ƒ/1.8) set on the microscope. Following ages without nematodes, distributed in the same 70-20-10 each test, the device is washed and cleaned in accordance ratio. The CNN was trained with a batch size of 8, over with the previously described procedures. 200 epochs, and an image size of 1920x1080 to enhance data quality when dealing with small objects. Afterward, 2.5. Deep learning detection 32 one-minute inference videos were analyzed, evenly split between stimulus and control contexts. Frames were A deep learning approach was utilized to diferentiate extracted from each video at a frequency of 1Hz. For each between scenarios involving stimulus presence and con- frame, the CNN detected nematodes distinguished under trol conditions for nematodes. Beyond discriminating stimulus and control conditions. The outcomes were avbetween the two contexts, this approach enabled the eraged for each frame to establish the presence or absence analysis of motor activity diferences captured through of a stimulus in the scenario. This procedure was iterated images, attributed to the stimulus presence. In this sense, for each video. In evaluating the CNN model, various nematodes have been employed as biosensors to gather metrics such as accuracy, precision, recall, f1-score, and information about the surrounding environment, while mAP were considered. Training and subsequent analyses also investigating changes in behavioral traits in response were performed using Ultralytics in Python with a Tesla to a stimulus using deep learning techniques. To achieve T4 GPU. this objective, a YOLOv8n Convolutional Neural Network (CNN) model pre-trained on the COCO dataset, a widely 2.6. Optical flow analysis used benchmark for object detection tasks [14], was utilized. The selection of this CNN model was driven by its For a comprehensive motor activity investigation, an opversatility and demonstrated efectiveness in object detec- tical flow analysis was integrated into the deep learning tion tasks. YOLO networks have been applied across vari- framework, using the Farnebäck algorithm [19] with speous detection tasks, particularly in the v8 version, encom- cific parameter settings (image scale of 0.5, 5 pyramid passing small object identification in Unmanned Aerial layers, averaging window size of 15, 3 iterations, size of Vehicle (UAV) images [15], evaluating medical face mask the pixel neighborhood used for polynomial expansion adherence in COVID-19 scenarios [16], or also detecting in each pixel of 7, standard deviation of the Gaussian diverse marine species [17]. YOLOv8 employs a CNN used to smooth derivatives for the polynomial expansion structured into three main components: the backbone, of 1.5). A total of 60 videos, balanced between control neck, and head. The backbone, based on a modified CSP- and stimulus contexts, were considered. Given that neDarknet53 architecture, facilitates multi-scaled object de- matodes exhibit body movements with a 2Hz frequency tection through a feature pyramid network, generating according to literature [20], frames were sampled at 5Hz ifve scale features [ 18]. The neck introduces a PAN-FPN in accordance with the Shannon theorem for the optical architecture, optimizing performance while maintain- flow analysis. First, the optical flow magnitude was coming a lightweight design. The detection head consists of puted by comparing two frames. The CNN model was convolutional and fully connected layers for object classi- exploited to obtain the bounding boxes of the detected ifcation and bounding box regression. Loss functions like nematodes. Then, the mean magnitude value for each BCE Loss and DFL/CIoU are employed. YOLOv8 operates bounding box was considered. Lastly, the maximum magas an anchor-free detection model, directly predicting nitude among the averaged values of the bounding boxes object centers. Dynamic sample assignment is achieved was extracted for each frame, to avoid misclassifications through the Task-Aligned Assigner mechanism, further with background. Temporal data were collected, and the improving accuracy and robustness. Architectural en- mean of these values was computed to derive a represenhancements include module exchanges and convolution tative value for each video. The analysis was conducted replacements, resulting in a model weighing 6.24 MB using OpenCV in Python. in the v8 nano version. The pre-trained CNN model was fine-tuned to identify and distinguish nematodes in 2.7. Statistical analysis the presence of a stimulus from those in a control setting. A dataset of 1680 images was obtained from 50 Data were normally distributed (Shapiro-Wilk test, videos, evenly distributed between control and stimulus p>0.01) and homoscedastic (Levene test, p>0.01). Therecontexts. The dataset was divided into training (70%), val- fore, statistical significance between stimuli and control idation (20%), and testing (10%) sets, resulting in subsets was established using a t-test. Statistical analyses were comprising 1189, 339, and 152 images, respectively. An performed using JMP Pro 17 software. The threshold was equal distribution of images across both classes was en- set at p=0.05.
identifying the control context, as demonstrated in the normalized confusion matrix in Fig.2(C). Examination of Microscopical observations have allowed us to identify the normalized confusion matrix indicates that the conthe main behaviors performed by S. carpocapsae, includ- trol class achieved 100% accuracy, while the stimulus class ing an increase in movement and the size of oscillations. exhibited 88% accuracy with a 12% error attributed to misThis clearly demonstrates the presence of an efect due classification as the control class. Fig. 2( D) showcases to tested cue arising from the host on the gestural com- predictions in video analysis. Overall, the video analysis plex of the EPN. Our findings indicate that S. carpocapsae attained an accuracy of 0.938, precision of 1, recall of exhibits not only a sit-and-wait strategy but also demon- 0.875, and f1-score of 0.933 in identifying stimulus presstrates active host-seeking behavior in response to chem- ence from behavioral traits. The results obtained through ical stimuli within its proximity. The fact that nematodes deep learning validate the observations. The CNN efecchanged their posture and locomotion direction towards tively distinguished between the two contexts (stimulus the stimulus is functional to describe L. botrana feces as presence and control) based on nematodes. Considering an attractive cue. Furthermore, this characteristic has this image-based approach, the diferences are attributed been sparsely evaluated in the literature [21] but holds to altered gestures due to the presence of a stimulus. The significant value for enhancing our understanding about subsequent analysis integrating the optical flow aims this species ethology and optimizing its utilization as to investigate changes in motor activity over time beBCA. The CNN model achieved a precision of 0.632, re- tween the control and stimulus contexts. Fig.3(A) shows call of 0.623, and mAP0.5 of 0.608 on the validation set. a graphical representation of the magnitude for the stimPredictions for two images outside the training and vali- ulus scenario, obtained through MINMAX normalization. dation processes are shown in Fig.2(A) and (B). Although Initially, the CNN model was utilized to obtain bounding the network exhibits limitations, particularly in individ- boxes for each frame. Subsequently, the mean optical ual identification compared to background, conducting lfow magnitude was calculated for each bounding box. video analysis yields promising results with no errors in Finally, the maximum value among the averaged mag
nitude values in the bounding boxes was determined. not to the emitter. Nematode responses were assessed
Temporal profiles of this value for the control and stim- using automated motility tracking and computer vision
ulus contexts are compared in Fig.3(B) considering two techniques [24]. The proposed integrated methodology
video samples. The mean of these magnitude values over significantly enhances our ability to discern and
undertime was considered to obtain an overall value for each stand EPN responses to diverse host stimuli. The use
video. Fig. 3(C) illustrates the comparison between con- of AI-based techniques and deep learning, integrated
trol and stimulus data. Statistical analysis conducted on with microfluidics and optical flow analysis, represents
this data demonstrates significance ( p=0.010), providing an innovative approach to behavioral studies, ofering
evidence of observed diferences in motor activity. The the possibility of using nematodes as biosensors.
Furintegration of optical flow with deep learning detection ther research on EPN behavior, utilizing pose estimation
reveals increased motor activity of nematodes in the pres- and diverse organic stimuli, could enhance commercial
ence of the stimulus when compared to the motility of product performance and expand S. carpocapsae’s use
control specimens. as a BCA for other phytophagous species, supporting
biological control principles [
The deep learning approach efectively diferentiated between scenarios with stimuli and control conditions for This research was supported by the EU H2020 FETOPEN nematodes. Beyond discrimination, it helped analyze Project “Robocoenosis - ROBOts in cooperation with a motor activity the presence of the stimulus. EPNs have bioCOENOSIS” [899520], the PRIN Project “COSMIC been used as biosensors, gathering environmental data COntrolled Space MIcroecological system supporting and revealing behavioral changes through deep learning eCopoiesis” granted by the Italian Ministry of Educatechniques. We utilized various methods, including mi- tion, University and Research (MIUR) [2022EY5BXC], and crofluidics, deep learning, and optical flow, to investigate the EU H2020-MSCARISE-2018 “ECOBOTICS.SEA - Biothese diferences and develop an automated workflow inspired Technologies for a Sustainable Marine Ecosysfor studying EPNs. Microfluidics was already employed tem” [824043]. Funders had no role in the study design, in several biological and behavioral assays with diferent data collection and analysis, decision to publish, or prepaspecies [22], but considering nematodes, this technique ration of the manuscript. The authors are grateful to Ms. is common also for the model organism Caenorhabdi- Cristina Piras for her kind support in visualization. tis elegans [23]. Our innovative approach is essential for studying a parasite’s behavior and locomotion in a natural-like condition. AI analysis revealed challenges References at the frame level due to background interference, but exploring diferent architectures may improve results.
However, at the video level, the model showed satisfactory performance with 100% precision. Optical flow and statistical analyses highlighted distinctions in motor activity between individuals in diferent contexts. Tests on EPN movement emphasized their ability to sense, move towards, and enter hosts. AI-based results confirm the preliminary microscopic observations regarding the presence of active motility in the tested species and supports the few assays which consider S. carpocapsae a nematode with some cruiser’s features [21]. Considering our results it’s clearly observed that S. carpocapsae has an actual locomotion towards the stimulus, explainable as chemotaxis where there is a spatial movement and klinotaxis where speciemens change their body conformation by bending or turning in presence of the organic stimulus.
Nematodes demonstrate aggregate migration; notably, we observed increased presence near the attractor [23].
According to our results, we are able to define larval feces cues as a kairomone; in fact, they evoke a physiological response in EPNs, which is favorable to the receiver but