Automated Detection of Uninformative Frames in Pulmonary Optical Endomicroscopy

Abstract

The detection algorithm is accurate and robust in pulmonary OEM frame sequences. Conditional to appropriate model refinement, the algorithms can become applicable in other organs.

Transfer learning for informative-frame selection in laryngoscopic videos through learned features

“…A statistical approach to informative-frame selection in esophageal microscopy images, which exploits intensity, entropy and keypoint-based features, is proposed in [23]. Texture-based features from lung microscopy images are classified with Gaussian mixture models in [12]. In [11], a set of intensity, keypoint-based and textural features and multi-class SVMs are used to classify informative and three classes of uninformative frames in laryngoscopic videos in narrow-band imaging (NBI).…”

“…However, this operation is qualitative, prone to human error and usually time consuming [10]. A reasonable alternative to visual assessment is the automatic selection of informative frames, which is however not always trivial due to variability in image characteristics (e.g., noise level and resolution), image acquisition protocols, and tissue anatomy [11,12].…”

Enhanced avidity from a multivalent fluorescent antimicrobial peptide enables pathogen detection in a human lung model

“…This approach shows particular promise for the future as it can be multiplexed using fiber based systems capable of multispectral imaging 24,25 , used in combination with other SmartProbes for Gram-negative bacteria 13 or inflammatory cells 26,27 and compounds can be delivered directly into the imaging field of view 28 . Furthermore, with ongoing refinements to the image analysis algorithms through multiple methods 29–31 , automated readouts of the signals generated may help decision making and advance our understanding of the pathophysiology during suspected pneumonia.…”

Image computing for fibre-bundle endomicroscopy: A review

“…Feature spaces combining two or more of the above descriptors are also frequent, with descriptors customarily extracted from the whole image, yet in some cases, regular or randomly distributed sub-windows/patches have been used, either on their own, or in conjunction to the whole image feature space. A number of well-established classifiers have been assessed, including (i) k-Nearest Neighbours (kNN) (André et al, 2012b;Desir et al, 2010;Hebert et al, 2012;Saint-Réquier et al, 2009;Srivastava et al, 2005;Srivastava et al, 2008), (ii) Linear and Quadratic Discriminant Analysis (LDA and QDA) (Leonovych et al, 2018;Srivastava et al, 2005;Srivastava et al, 2008), (iii) Support Vector Machines (SVM) and their adaptation with Recursive Feature Elimination (SVM-RFE) (Desir et al, 2010;Desir et al, 2012b;Jaremenko et al, 2015;Leonovych et al, 2018;Petitjean et al, 2009;Rakotomamonjy et al, 2014;Saint-Réquier et al, 2009;Vo et al, 2017;Wan et al, 2015;Zubiolo et al, 2014), (iv) Random Forests (RF) and variants such as Extremely Randomised Trees (ET) (Desir et al, 2012a;Heutte et al, 2016;Jaremenko et al, 2015;Leonovych et al, 2018;Seth et al, 2016;Vo et al, 2017), (v) Gaussian Mixture Models (GMM) (He et al, 2012;Perperidis et al, 2016), (vi) Boosted Cascade of Classifiers (Hebert et al, 2012), (vii) Neural Networks (NN) (Ştefănescu et al, 2016), (viii) Gaussian Processes Classifiers (GPC), and (ix) Lasso Generalised Linear Models (GLM) (Seth et al, 2016). Most studies employed leave-k-out and k-fold cross validation to assess the predictive capacity of the proposed methodology on limited, pre-annotated frames.…”