Real-time monitoring of the status of individual living organisms by image analysis

Effective management of living organisms requires acquiring reliable information of the system of interest. Cameras, in combination with image analysis techniques, can be used to measure multiple information from the living organism in a non-intrusive, cheap and reliable way by reducing the redundant information of the image to a set of representative feature variables that retain the relevant information of the process. Therefore, selecting the proper feature variables is essential in developing monitoring systems that aim to gain information that supports and optimises decision making or to function as an early warning system. A feature variable does not necessarily have to be complex, but it needs to be valid and reliable so that it truly measures what is intended to be measured over a series of time, places and individuals. The feature variable should also be specific and sensitive so that only few false positives and false negatives are detected. A feature variable needs also to be measurable at a reasonable cost and effort. Only few variables, but measured continuously can be used to monitor the status or to detect changes from the normal status of living organisms. Taking into consideration the properties of the feature variables, the essential connection between the domain of biology and technology in context of selecting the proper feature variables becomes evident. Only biology experts have the insight into complex systems such as living organisms and can provide a pool of possible indicators appropriate for the specific process. However, not every indicator can be measured in a reliable and robust way and can become a feature variable. Therefore, the choice among these indicators must be technology driven due to the limitation of available technology. Furthermore, monitoring of living organisms in particular poses a challenge because they are Complex, Individually different, Time-varying and Dynamic ‘CITD’ systems. Since every living organism is individually different from each other, no single organism responds like the average of a group. And since living organisms are time- varying, a single organism’s response to a stimulus might be different each time the stimulus is applied. Monitoring systems therefore have to cope with a variation both among different individuals (individuality) and within the same individual (time- variance). The objectives of this thesis were firstly to develop monitoring systems for living organisms by selecting few but appropriate feature variables, secondly to compare population-based and individual models for the monitoring of living organisms and thirdly to develop on-line and real-time monitoring systems that allow not only to detect but to predict the status of living organisms. In chapter 2, image analysis was used to estimate accurately the weight of Jade perch Scortum barcoo. Different feature variables were tested. Measuring the area of the fish body, excluding the tail fins, from side view, proved to be the most robust and accurate to predict the fish weight, resulting in a coefficient of determination (R2) of 0.99 and a mean absolute relative error (MARE) of 6% when compared to the manual weight measurement. In chapter 3, it was demonstrated how feature variables extracted by image analysis and measured over time can be used to detect changes in the normal walking behaviour in dairy cows and therefore to detect lameness. Cows respond to lameness by changing their gait in order to avoid bearing weight on the injured limb. Image analysis was therefore used to measure the cow’s posture while walking by using only four feature variables describing the geometry of the back spine. Based on the average population information, these feature variables were then used by a classification algorithm in order to discriminate between lame and not lame cows. In chapter 4, it was then illustrated that cows’ responses to lameness are different compared to the average of the group. The population-based approach of chapter 3, in fact, did not take into account the high inter-variability among cows and its impact on cow locomotion. A population-based approach maximizes the accuracy of the average of the population only theoretically because almost no cow fulfils the specifications of a population average in reality. In other words, if one cow is not within a certain standard deviation of its scoring group, it will be misclassified. In chapter 4, an individual approach was used in order to compensate this inaccuracy by considering the high inter-variability among cows. It was shown that the accuracy in detecting lameness on a three-level scale (Not Lame, Mildly Lame, Severely Lame) increased by 10% up to 91% when the individual-based approach was used. In chapter 5, feature variables related to the motion of interacting pigs were used in order to detect aggressive interactions among animals. An image whose pixel intensity was a function of motion was then introduced. This image allowed extracting feature variables about both the spatial region in which the motion took place and the evolvement of the motion. Feature variables from this image were calculated over time, thus obtaining time series of feature variables. Based on the average population information, statistical information was then used in order to detect aggressive interactions with an accuracy of 89%. In chapter 6, it was demonstrated that even cells, as a relatively simple organism, react differently to the same enzymatic solution even if applied under the same conditions. Therefore an average time to decide when to stop the enzymatic reaction and start the mechanical stimulation to detach the cell from the plate should not be used. Data-based modelling techniques were used to measure the cell’s morphological response to the harvesting solution (R2 = 0.99±0.005) and to calculate the optimal time to harvest and therefore to avoid under or over exposure of the individual cell to the enzymatic solution. Furthermore, by on-line monitoring cell morphology it was possible to predict the harvesting time-point as soon as the model parameters converged with a prediction error of 23±23 (Mean±SE) seconds and therefore to help the user to make a decision based on objective measurements, resulting in an increased viability and homogeneity of cells.status: publishe