Synapse classification and localization in Electron Micrographs

Classification and detection of biological structures in Electron Micrographs (EM) is a relatively new large scale image analysis problem. The primary challenges are in modeling diverse visual characteristics and development of scalable techniques. In this paper we propose novel methods for synapse detection and localization, an important problem in connectomics. We first propose an attribute based descriptor for characterizing synaptic junctions. These descriptors are task specific, low dimensional and can be scaled across large image sizes. Subsequently, techniques for fast localization of these junctions are proposed. Experimental results on images acquired from a mammalian retinal tissue compare favorably with state of the art descriptors used for object detection. Visual classification of structures of interest has a wide variety of applications in natural images, video sequences, aerial and biological images. At one end of the spectrum, detection and classification of objects in natural images has received a significant research interest in recent times with competitions like PASCAL and ImageNet. Algorithms such as the DPM and Sparselet have been shown to perform extremely well on such challenges. At the other end of the spectrum are emerging applications in bio-microscopic imagery, where automated image analysis is crucial due to high throughput image acquisition. Constructing an over-arching classification/detection model that can work across any bio-microscopic imagery is challenging due to inherent variability in imaging protocols. For instance, a tissue imaged using different imaging conditions, such as the light, confocal or electron micros-copy, can lead to visually very different images. Knowledge of associated meta-data such as molecule specific bio-markers used for imaging are critical for further processing and interpretation of such images. As a result, an algorithm developed for one modality is difficult to adopt to another modality, necessitating the development of application specific classification/detection algorithms. The scope and applications to problems in bio-microscopic imagery are fairly diverse, with many applications still relatively unexplored. We focus on one such application, namely structural connectomics. Connectomics: Connectomics is a sub-field of neuroscience aiming to understand neuronal circuitry in the animal brain. Synapses, or edges in the neuronal circuit graph can be resolved only at nano-meter (10 À9 m) resolutions. Such resolutions require the acquisition of massive amounts of data, typically ranging into several terabytes. Due to recent developments in high throughput micros-copy, such datasets can be acquired in a fully automated fashion without any human intervention. The bottleneck manifests in analyzing these large image mosaics, which could …