Automated classification and tracking of microscopic holographic patterns of nematodes using machine learning methods

Analysing nematode behaviour helps estimate biomechanical parameters for applications like cellular biology, pharmacology and cognitive neuroscience. Portable holographic platforms offer cost-effective, high-resolution, high-frame-rate, wide-field imaging compared to conventional microscopy. Holographic methods can reconstruct original shapes using numerical diffraction, although this is computationally expensive. However, video holography remains challenging due to the fast motion and overlapping of holograms when nematodes swim in crowded environments. In this work we address this problem by focusing on automated detection and tracking of nematodes in densely populated environments, using machine learning methods. The main advantage of our approach is to present an automated computational flow to detect and analyse the behaviour of live nematodes in video directly from the raw holographic signals, without the requirement of phase-recovering methods for diffraction. For this purpose, we developed a three-step CNN-based approach consisting of: i) nematode hologram detection; ii) temporal tracking; and iii) behavioural analysis based on mobility parameters. In terms of precision, the obtained results show that using a two-stage detector, the Faster R-CNN architecture with the ResNet18 model as a backbone, presented the best Mean Average Precision (mAP) score with 86% for correct classification. For tracking, the best performing algorithm was IoU with a HOTA, resulting in 62.42% when applied on the individually tagged nematodes, which is comparable to the best current generic multi-tracking approaches available over the literature. Our obtained results show that the use of a convolutional neural network approach associated with a classic tracking algorithm is a very suitable approach for nematode detection and behavioural analysis for biological assays directly from holograms, even in densely populated environments. The proposed approach has been presented as a promising solution for automated inspection of free-living nematode individuals.