Bayesian image restoration and bacteria detection in optical endomicroscopy
Karam Mohammed AbdElkarim, Ahmed
MetadataShow full item record
Optical microscopy systems can be used to obtain high-resolution microscopic images of tissue cultures and ex vivo tissue samples. This imaging technique can be translated for in vivo, in situ applications by using optical ﬁbres and miniature optics. Fibred optical endomicroscopy (OEM) can enable optical biopsy in organs inaccessible by any other imaging systems, and hence can provide rapid and accurate diagnosis in a short time. The raw data the system produce is diﬃcult to interpret as it is modulated by a ﬁbre bundle pattern, producing what is called the “honeycomb eﬀect”. Moreover, the data is further degraded due to the ﬁbre core cross coupling problem. On the other hand, there is an unmet clinical need for automatic tools that can help the clinicians to detect ﬂuorescently labelled bacteria in distal lung images. The aim of this thesis is to develop advanced image processing algorithms that can address the above mentioned problems. First, we provide a statistical model for the ﬁbre core cross coupling problem and the sparse sampling by imaging ﬁbre bundles (honeycomb artefact), which are formulated here as a restoration problem for the ﬁrst time in the literature. We then introduce a non-linear interpolation method, based on Gaussian processes regression, in order to recover an interpretable scene from the deconvolved data. Second, we develop two bacteria detection algorithms, each of which provides diﬀerent characteristics. The ﬁrst approach considers joint formulation to the sparse coding and anomaly detection problems. The anomalies here are considered as candidate bacteria, which are annotated with the help of a trained clinician. Although this approach provides good detection performance and outperforms existing methods in the literature, the user has to carefully tune some crucial model parameters. Hence, we propose a more adaptive approach, for which a Bayesian framework is adopted. This approach not only outperforms the proposed supervised approach and existing methods in the literature but also provides computation time that competes with optimization-based methods.