Fig 1.
Overview of biological structure detection using multi-tiered classification.
a) Unsupervised image processing techniques are often necessary to harness the power of emerging imaging and experimental technologies. b) An overview of the proposed generalizable two layer classification architecture for the autonomous identification of specific biological structures. Intrinsic, computationally simple features and relational or computationally expensive features are partitioned into two layers to accommodate both structural complexity and efficiency.
Fig 2.
Preprocessing and feature selection for head versus tail discrimination in C. elegans.
a) The limited field of view of high resolution imaging systems creates a need for spatial positioning along the anterior-posterior axis of the worm. As a landmark for orienting the A-P axis, the head of the worm is distinguished by the presence of the pharynx and a grinder structure (inset below). b) Preprocessing for bright field structural detection consists of minimum intensity projection of a sparse z-stack (MP) followed by Niblack local thresholding (BW0) and preliminary filtration of segmented particles to generate candidates for subsequent classification (BW1). c) In layer 1 of classification, computationally inexpensive, intrinsic properties of the candidates (BW1) are calculated for SVM classification and reduction of the candidate pool (BW2). d) Two example image processing sequences showing that while the shape-intrinsic features used in layer 1 of classification significantly reduces the candidate pool, it is insufficient for robust, specific identification of the grinder particle. e) From the reduced candidate pool, layer 2 of classification utilizes regional properties of the remaining candidates to distinguish the grinder from other structural and textural elements of the worm body with high specificity, making identification of the head possible on the basis of the presence of the grinder particle.
Fig 3.
Optimization and training of the two layers of SVM classification for pharyngeal grinder detection.
a) To construct the layer 1 classifier with the specified feature set, five-fold cross-validation with a manually annotated training set is first used to optimize SVM model parameters and ensure classification performance. b) Classification performance based on the false positive (FPR) and false negative (FNR) error rates observed in five-fold cross-validation allows selection of an optimal parameter set. c) The full training set and optimized parameters are used to construct the final layer 1 SVM model. Linear projections of the training set features onto two dimensions show that the layer 1 feature set and the optimized SVM model are insufficient for identifying the grinder particle with high specificity. d) The second layer of classification refines the final classification decision and is parameter-optimized using the candidates passed from layer 1 of classification. e) Classification performance based on five-fold cross-validation is used for parameter selection. f) The reduced layer 2 training set and optimized parameters are used to construct the final layer 2 SVM model. Linear projections of layer 2 features for the training set demonstrate the capability of a two layer scheme for the detection of the grinder with both high specificity and sensitivity.
Fig 4.
Head versus tail classification using grinder detection is robust to changes in experimental conditions and genetic background.
a) Changes in experimental conditions, such as food availability, can alter the bulk morphology and the appearance of worm body in bright field, with potential consequences for classification accuracy. b) Our head versus tail classification scheme maintains sensitivity and specificity at over 95% at different ages and feeding conditions despite these biological changes. c) Genetic changes can also induce changes in bulk morphology and texture of the worm. d) Despite not being represented within the training set, the performance of the classifier is maintained even for mutant worms (dpy-4 (-)) with major morphological changes. e) Changes in the optics, camera or acquisition parameters can alter the final resolution of images. f) The inclusion of the calibration metric within feature calculation (S2 Fig and S3 Fig) maintains classifier performance across a two-fold change in image resolution due to alternations in digital binning.
Fig 5.
First layer classification for detection of fluorescently labelled neuronal cells demonstrates generalizability of first layer features for particle shape classification.
a) Stereotypical positioning of the ASI neuron pair in the head of the worm. Many neuronal cells in the worm are organized as similar pairs near the pharynx. b) Bright field and fluorescent maximum intensity projection showing the appearance and positioning of fluorescently labelled ASI cells in the head of the worm. c) Preprocessing of raw fluorescent images showing binary image after Niblack thresholding (BW0) and initial filtration of the candidate set by size (BW1). d) First layer classification of fluorescently labeled neurons shows good generalizability of the first layer feature set developed for pharyngeal grinder detection for classification based on binary particle shape.
Fig 6.
Second layer classification for neuron pair detection.
a) The first layer of classification is insufficient for rejection of all background particles. b) The reduced candidate set from the first layer of classification is used to form candidate cell pairs with feature sets describing their relative positioning and intensities. c) Although classification based on these features is sufficient for accurate cell pair detection in the majority of cases (left), multiple potential cell pairs are sometimes classified within the same image (right). d) Incorporating probability estimates (shown in panel c) into the SVM model and selecting the most likely cell pair eliminates these false positives and increases the specificity of the classifier.
Fig 7.
Second layer classifier for cell pattern recognition and identification.
a) Representative maximum intensity projection and schematic representation of the two neuron pairs in which an insulin-like peptide is expressed. b) The modularity of our scheme permits the preprocessing and layer 1 classification components from neuron pair detection to be re-used for the recognition and identification of these neuron pairs. c) To identify the pattern with the appropriate cell identifications, properties for all possible combinations and arrangements of the layer 1 candidates are calculated. Here, all six such candidate sets for 4 candidate particles are shown. d) Validation of the SVM classifier trained with these features shows high specificity but only moderate sensitivity. e) The lower sensitivity observed for this classification scheme is mainly due to the limit ability to accommodate biological deviations from the stereotypical arrangement of the neurons while still maintaining high specificity.