Automating the study of cell motility opens the opportunity to accelerate the analysis for treatment of several illnesses caused by cancer cells, evaluating the performance of new drugs and repositioning existing drugs for customised medication. Differently from other brain cells, the glioblastoma stem cells have greater motility, and some drugs can boost their expansion.
The challenge of tracking and measuring their motility comes as a formidable challenge for most of the computer vision techniques, with low-quality images, segmentation issues, instance splitting and shape variance. This work explores the utilisation of a two-stage pipeline composed by a feedbacked Otsu’s threshold-based detection and Kernelised Correlated Filters object tracking for detecting and carrying out the object tracking part of the problem. Additionally, the labelling process utilises a weighted multi-feature approach, which takes several features from the objects and matches them for tracking recovery with the aim of label preservation. By utilising this approach, the detection error obtained was around 17%.
In contrast, the tracking valid ratio (healthy trackers over detections) achieved 100% for a single-scene approach, and 77.11% for a multi-scene approach, for a non-chaotic and well-behaved video sequence. Additionally, this work presents a performance optimisation for a collimation algorithm. The optimisation consists on data-layout transformations, AVX-512 vectorisation, and multithreading applied to a perforated loop. At the end of the work, the optimisation manages to get 4.65 times of speed-up by using six threads, and AVX-512 intrinsics.