Sri Phani Krishna Karri

Transfer learning based classification of optical coherence tomography images with diabetic macular edema and dry age-related macular degeneration

We present an algorithm for identifying retinal pathologies given retinal optical coherence tomography (OCT) images. Our approach fine-tunes a pre-trained convolutional neural network (CNN), GoogLeNet, to improve its prediction capability (compared to random initialization training) and identifies salient responses during prediction to understand learned filter characteristics. We considered a data set containing subjects with diabetic macular edema, or dry age-related macular degeneration, or no pathology. The fine-tuned CNN could effectively identify pathologies in comparison to classical learning. Our algorithm aims to demonstrate that models trained on non-medical images can be fine-tuned for classifying OCT images with limited training data.

ReLayNet: retinal layer and fluid segmentation of macular optical coherence tomography using fully convolutional networks

Optical coherence tomography (OCT) is used for non-invasive diagnosis of diabetic macular edema assessing the retinal layers. In this paper, we propose a new fully convolutional deep architecture, termed ReLayNet, for end-to-end segmentation of retinal layers and fluid masses in eye OCT scans. ReLayNet uses a contracting path of convolutional blocks (encoders) to learn a hierarchy of contextual features, followed by an expansive path of convolutional blocks (decoders) for semantic segmentation. ReLayNet is trained to optimize a joint loss function comprising of weighted logistic regression and Dice overlap loss. The framework is validated on a publicly available benchmark dataset with comparisons against five state-of-the-art segmentation methods including two deep learning based approaches to substantiate its effectiveness.

Learning layer-specific edges for segmenting retinal layers with large deformations.

We present an algorithm for layer-specific edge detection in retinal optical coherence tomography images through a structured learning algorithm to reinforce traditional graph-based retinal layer segmentation. The proposed algorithm simultaneously identifies individual layers and their corresponding edges, resulting in the computation of layer-specific edges in 1 second. These edges augment classical dynamic programming based segmentation under layer deformation, shadow artifacts noise, and without heuristics or prior knowledge. We considered Dukes online data set containing 110 B-scans of 10 diabetic macular edema subjects with 8 retinal layers annotated by two experts for experimentation, and achieved a mean distance error of 1.38 pixels whereas that of the state-of-the-art was 1.68 pixels.

Detection of retinal vessels in fundus images through transfer learning of tissue specific photon interaction statistical physics

Loss of visual acuity on account of retina-related vision impairment can be partly prevented through periodic screening with fundus color imaging. Largescale screening is currently challenged by inability to exhaustively detect fine blood vessels crucial to disease diagnosis. In this work we present a framework for reliable blood vessel detection in fundus color imaging through inductive transfer learning of photon-tissue interaction statistical physics. The source task estimates photon-tissue interaction as a spatially localized Poisson process of photons sensed by the RGB sensor. The target task identifies vascular and non-vascular tissues using knowledge transferred from source task. The source and target domains are retinal images obtained using a color fundus camera with white-light illumination. In experimental evaluation with the DRIVE database, we achieve the objective of vessel detection with max. avg. accuracy of 0.9766 and kappa of 0.8213.

In situ histology of mice skin through transfer learning of tissue energy interaction in optical coherence tomography

Abstract. Tissue characterization method in optical coherence tomography (OCT) for in situ histology of soft tissues is presented and demonstrated for mice skin. OCT allows direct noninvasive visualization of subsurface anatomy. It is currently used for in situ investigation of lesions in skin, vessels, retinal layers, oral, and bronchial cavitities. Although OCT images present high resolution information about tissue morphology, reporting requires a reader experienced in interpretation of the images, viz., identification of anatomical layers and structures constituting the organ based on OCT speckle appearance. Our approach characterizes tissues through transfer learning of tissue energy interaction statistical physics models of ballistic and near-ballistic photons. The clinical information yield with our approach is comparable to conventional invasive histology. On cross evaluation with a mice model experiment, the epidermis, papillary dermis, dermis, and adipose tissue constituting the mice skin are identified with an accuracy of 99%, 95%, 99%, and 98%, respectively. This high accuracy of characterizing heterogeneous tissues using OCT justifies the ability of our computational approach to perform in situ histology and can be extended to regular clinical practice for diagnosis of vascular, retinal, or oral pathologies.

Deep learning of tissue specific speckle representations in optical coherence tomography and deeper exploration for in situ histology

Optical coherence tomography (OCT) relies on speckle image formation by coherent sensing of photons diffracted from a broadband laser source incident on tissues. Its non-ionizing nature and tissue specific speckle appearance has leveraged rapid clinical translation for non-invasive high-resolution in situ imaging of critical organs and tissue viz. coronary vessels, healing wounds, retina and choroid. However the stochastic nature of speckles introduces inter- and intra-observer reporting variability challenges. This paper proposes a deep neural network (DNN) based architecture for unsupervised learning of speckle representations in swept-source OCT using denoising auto-encoders (DAE) and supervised learning of tissue specifics using stacked DAEs for histologically characterizing healthy skin and healing wounds with the aim of reducing clinical reporting variability. Performance of our deep learning based tissue characterization method in comparison with conventional histology of healthy and wounded mice skin strongly advocates its use for in situ histology of live tissues.

Classifying histopathology whole-slides using fusion of decisions from deep convolutional network on a collection of random multi-views at multi-magnification

Histopathology forms the gold standard for confirmed diagnosis of a suspicious hyperplasia being benign or malignant and for its sub-typing. While techniques like whole-slide imaging have enabled computer assisted analysis for exhaustive reporting of the tissue section, it has also given rise to the big-data deluge and the time complexity associated with processing GBs of image data acquired over multiple magnifications. Since preliminary screening of a slide into benign or malignant carried out on the fly during the digitization process can reduce a Pathologists work load, to devote more time for detailed analysis, slide screening has to be performed on the fly with high sensitivity. We propose a deep convolutional neural network (CNN) based solution, where we analyse images from random number of regions of the tissue section at multiple magnifications without any necessity of view correspondence across magnifications. Further a majority voting based approach is used for slide level diagnosis, i.e., the class posteriori estimate of each views at a particular magnification is obtained from the magnification specific CNN, and subsequently posteriori estimate across random multi-views at multi-magnification are voting filtered to provide a slide level diagnosis. We have experimentally evaluated performance using a patient level 5-folded cross-validation with 58 malignant and 24 benign cases of breast tumors to obtain average accuracy of 94.67 ± 14.60%, sensitivity of 96.00 ± 8.94%, specificity of 92.00 ± 17.85% and F-score of 96.24 ± 5.29% while processing each view in ≈ 10 ms.

Transfer learning of tissue photon interaction in optical coherence tomography towardsin vivo histology of the oral mucosa.

Oral cancer evolves from different premalignant conditions and the key to save lives is through diagnosis of early symptoms. The conventional practice of post biopsy histopathology reporting is dependent on specificity of sampling site and optical coherence tomography (OCT) imaging is clinically used for guidance. Clinicians infer the tissue constitution by interpreting intensity images and are challenged by inter-and intra-observer variability. In this paper we propose transfer learning of tissue specific photon interaction statistical physics in swept-source OCT for characterizing the oral mucosa with the aim of reducing this reporting variability. The source task models statistical physics of ballistic and near-ballistic photons and its intensity attenuation and target task learns the parameters obtained by solving the source task to identify co-located heterogeneity of tissues. Performance is compared with conventional histopathology of healthy, premalignant and malignant oral lesions supporting its use towards in vivo histology of the oral mucosa for pre-biopsy screening.

Computer assisted cervical cytological nucleus localization

Cervical cancer is the second most common form of malignancy among women in India. Regular screening of cervix can mollify its incidence thus enabling adoption of better prevention strategies. Proper localization and delineation of nuclear attributes for identification of crucial cellular features in Papanicolau stained cervical liquid based cytological images are tricky. Reduction of false negativity in the screening technique is the need of the hour. These ambiguities in the detection of nuclei are due to relative differences in staining intensity, presence of inflammatory cells, necrotic background, presence of bacteria, cellular overlapping causing super-imposition of nuclei, and clustering/clumping of cells etc. In this paper we present here a computationally lightweight yet elegant computer assisted automated technique for localization of epithelial cell nuclei in optical microscopic images of Pap stained monolayer cervical smears. The set of developed algorithms efficiently handle background separation and identification of nuclei in overlapping cells occurring as clusters. It uses morphological selection of region of interest preceded by intensity based object separation. The region of interest is iteratively bound using minimum bounding rectangles, to locate nuclei in the cell clusters. This method accurately inserts seed points, eliminates false seeds and detects nucleus. Thereafter region growing technique is applied considering obtained seed points to segment the nucleus from cells. It is inferred that most of the problems faced while locating nuclei are overcome with the above algorithm; the only cases when it fails is in presence of overlapping nuclei and in presence of overlapping neutrophils.

Learning scale-space representation of nucleus for accurate localization and segmentation of epithelial squamous nuclei in cervical smears

Computer vision systems are being introduced in pre-screening of cervical cytopathology slides to identify samples that require study by cytopathologists. These systems work on the principle of imaging and analysis of cytology features in general and nuclear features in particular. Thus accurate localization and segmentation of the nuclei is crucial for the systems. Though several methods have been conceptualized, developed and employed to achieve the tasks of localization and segmentation of nuclei in cytology images, most fail to localize nuclei with opened up chromatin. This paper presents a machine learning approach based framework for accurate localization and segmentation of nuclei. The approach uses the random forest model to learn complete scale-space representation of the nuclear chromatin distribution in green and color saturation channels. Based on the multi scale features of an unknown image this model can predict an image such that gray level value of a pixel is proportionate to the probability that the pixel belongs to nuclear region. This predicted image then can be used for accurate localization and segmentation of the nuclei by random walks approach. Accuracy of the system has been tested on a publicly available dataset images and was found to be approximately 97%.