Chabert, S. | bioRxiv Science

bioRxiv2026

PolliCrop: A high-throughput computer vision pipeline for pollinator monitoring in agroecosystems

Flower-visiting insect populations are declining since the 1990s, especially because of the decrease of floral resources in agricultural settings. Mass flowering crops can help increase resource availability, and plant breeding can be directed towards selecting varieties attracting more flower-visiting insects. This requires the implementation of an automated high-throughput phenotyping tool for assessing the attractiveness of plant genotypes to flower-visiting insects. In this study, (i) we present a procedure to take standardized images of sunflower heads with camera traps continuously at day and night in the field; (ii) we trained two versions of a deep learning model, named PolliCrop, to automatically detect and identify three classes of the main insects visiting sunflower on these images (non-Bombus bees, bumble bees, lepidopterans); (iii) we assessed and validated the ability of PolliCrop to correctly predict the true visitation frequencies of the insect classes on three sunflower genotypes; (iv) we presented two statistical approaches to compare the insect visitation frequencies between plant genotypes, one including weather variables, and the other one without. One PolliCrop version yielded satisfying performance to correctly detect the three insect classes. In particular, it correctly predicted the insect visitation frequencies on two sunflower genotypes in a range of {+/-}10%. The other PolliCrop version can be useful in certain contexts of images and objectives. PolliCrop can be extended in the future to other crop species by training PolliCrop on new images captured in these crops. The field experimental design to set up for comparing the attractiveness between genotypes is also discussed.

animal behavior and cognition↗

bioRxiv2026

PINN-ing the Balloon: A Physically Informed Neural Network Modelling the Nonlinear Haemodynamic Response Function in MRI

Accurate characterisation of the haemodynamic response function (HRF) is central to interpreting blood-oxygen-level-dependent (BOLD) signals in functional magnetic resonance imaging, yet standard estimation approaches remain centred around phenomenological formulations lacking biophysical grounding. We present a proof-of-concept methodological study: a physics-informed neural network (PINN) framework that bridges these paradigms by embedding the Balloon-Windkessel model directly into the training objective of a multi-headed neural network. Our approach simultaneously estimates probable latent neurovascular state variables such as cerebral blood inflow, metabolic rate of oxygen consumption, blood volume, and deoxyhaemoglobin content, through an indirect optimisation scheme in which the predicted BOLD signal is obtained via convolution of the estimated HRF with experimental stimuli. Training is governed by a composite loss, balancing differential-equation residuals, hard physics regularisation term, physiological initial conditions and data fidelity. In simulations with temporal signal-to-noise ratios representative of clinical acquisitions, the framework recovered ground-truth state variables with coefficients of determination exceeding 0.95 and mean squared errors below 10-3, at a physics-to-data weighting of 0.40:0.60. Application to 1.5 T block-design clinical data from an ischaemic stroke patient provides feasibility testing, yielding physiologically plausible, subject-specific HRF estimates, establishing feasibility of single-subject, physics-constrained HRF inference without reliance on fixed gamma basis assumptions. To our knowledge, this constitutes the first deployment of a single PINN incorporating the full Balloon-Windkessel model within an indirect training objective, reconstructing full BOLD observations, thereby positioning PINN-based haemodynamic modelling as a principled and personalised route towards more interpretable and patient-specific biomarkers.

neuroscience↗