Automatic detection of pathological myopia using machine learning

Pathological myopia is a severe case of myopia, i.e., nearsightedness. Pathological myopia is also known as degenerative myopia because it ultimately leads to blindness. In pathological myopia, certain myopia-specific pathologies occur at the eye’s posterior i.e., Foster-Fuchs’s spot, Cystoid degeneration, Liquefaction, Macular degeneration, Vitreous opacities, Weiss’s reflex, Posterior staphyloma, etc. This research is aimed at developing a machine learning (ML) approach for the automatic detection of pathological myopia based on fundus images. A deep learning technique of convolutional neural network (CNN) is employed for this purpose. A CNN model is developed in Spyder. The fundus images are first preprocessed. The preprocessed images are then fed to the designed CNN model. The CNN model automatically extracts the features from the input images and classifies the images i.e., normal image or pathological myopia. The best performing CNN model achieved an AUC score of 0.9845. The best validation loss obtained is 0.1457. The results show that the model can be successfully employed to detect pathological myopia from the fundus images.

Multi-Target Invisibly Trojaned Networks for Visual Recognition and Detection

Visual backdoor attack is a recently-emerging task which aims to implant trojans in a deep neural model. A trojaned model responds to a trojan-invoking trigger in a fully predictable manner while functioning normally otherwise. As a key motivating fact to this work, most triggers adopted in existing methods, such as a learned patterned block that overlays a benigh image, can be easily noticed by human. In this work, we take image recognition and detection as the demonstration tasks, building trojaned networks that are significantly less human-perceptible and can simultaneously attack multiple targets in an image. The main technical contributions are two-folds: first, under a relaxed attack mode, we formulate trigger embedding as an image steganography-and-steganalysis problem that conceals a secret image in another image in a decipherable and almost invisible way. In specific, a variable number of different triggers can be encoded into a same secret image and fed to an encoder module that does steganography. Secondly, we propose a generic split-and-merge scheme for training a trojaned model. Neurons are split into two sets, trained either for normal image recognition / detection or trojaning the model. To merge them, we novelly propose to hide trojan neurons within the nullspace of the normal ones, such that the two sets do not interfere with each other and the resultant model exhibits similar parameter statistics to a clean model. Comprehensive experiments are conducted on the datasets PASCAL VOC and Microsoft COCO (for detection) and a subset of ImageNet (for recognition). All results clearly demonstrate the effectiveness of our proposed visual trojan method.

Chief factors in Polish groups

In finite group theory, chief factors play an important and well-understood role in the structure theory. We here develop a theory of chief factors for Polish groups. In the development of this theory, we prove a version of the Schreier refinement theorem. We also prove a trichotomy for the structure of topologically characteristically simple Polish groups. The development of the theory of chief factors requires two independently interesting lines of study. First we consider injective, continuous homomorphisms with dense normal image. We show such maps admit a canonical factorisation via a semidirect product, and as a consequence, these maps preserve topological simplicity up to abelian error. We then define two generalisations of direct products and use these to isolate a notion of semisimplicity for Polish groups.

Examinations for creating an image of unacquired dose from the images of two types of dose in digital radiography

BACKGROUND: Digital radiography (DR) is grayscale adjustable and it can be unclear whether an acquired DR image is captured with the minimum radiation dose required. It is necessary to make an image of the amount of noise when taken at a lower dose than the acquired image, without increased exposure. OBJECTIVE: To examine whether an image of unacquired dose can be created from two types of dose DR images acquired using a phantom. METHODS: To create an additive image from two images of different doses, the pixel value of one image is multiplied by a coefficient and added to the other. The normalized noise power spectra (NNPS) of the normal image and the additive image with the same signal-to-noise ratio (SNR) are compared. The image noise of the unacquired doses is estimated from the graph changes of the pixel values and standard deviations of two images. The error between the SNR of the image obtained by changing the dose and the estimated SNR is measured. We propose a multiplication coefficient calculation formula that theoretically adjusts the additive image to the target SNR. The SNR error of the image created based on this formula is measured. RESULTS: The NNPS curves of the additive and normal images show a difference on the high frequency side. According to the statistics considering the preset of mAs value, there is no significant difference at 85%. The SNR estimation error is approximately 1%. The SNR error of the additive image created based on the formula is approximately 5%. CONCLUSION: The noise of the image of unacquired dose can be estimated, and the additive image adjusted to this value can be considered equivalent to the image taken at the actual dose.

Restoration and Calibration of Tilting Hyperspectral Super-Resolution Image

Tilting sampling is a novel sampling mode for achieving a higher resolution of hyperspectral imagery. However, most studies on the tilting image have only focused on a single band, which loses the features of hyperspectral imagery. This study focuses on the restoration of tilting hyperspectral imagery and the practicality of its results. First, we reduced the huge data of tilting hyperspectral imagery by the p-value sparse matrix band selection method (pSMBS). Then, we restored the reduced imagery by optimal reciprocal cell combined modulation transfer function (MTF) method. Next, we built the relationship between the restored tilting image and the original normal image. We employed the least square method to solve the calibration equation for each band. Finally, the calibrated tilting image and original normal image were both classified by the unsupervised classification method (K-means) to confirm the practicality of calibrated tilting images in remote sensing applications. The results of classification demonstrate the optimal reciprocal cell combined MTF method can effectively restore the tilting image and the calibrated tiling image can be used in remote sensing applications. The restored and calibrated tilting image has a higher resolution and better spectral fidelity.

Conditional Generative Adversarial Networks to Model iPSC-Derived Cancer Stem Cells

The realization of effective and low-cost drug discovery is imperative to enable people to easily purchase and use medicines when necessary. This paper reports a smart system for detecting iPSC-derived cancer stem cells by using conditional generative adversarial networks. This system with artificial intelligence (AI) accepts a normal image from a microscope and transforms it into a corresponding fluorescent-marked fake image. The AI system learns 10,221 sets of paired pictures as input. Consequently, the system’s performance shows that the correlation between true fluorescent-marked images and fake fluorescent-marked images is at most 0.80. This suggests the fundamental validity and feasibility of our proposed system. Moreover, this research opens a new way for AI-based drug discovery in the process of iPSC-derived cancer stem cell detection.

Overhaul of Three-D Image Based on Metamorphosis and Fabric Growing of Two-D Images

Shape is a critical physical property of normal and artificial three-D images that describes their outside appearances. Understanding differences among shapes and displaying the inconstancy inside and outside the shape classes are considered for shape analysis, and are the major issues in numerous applications, from normal image visualization to medical imaging. During diagnosis in medical image processing it is impossible to analyze the diseased areas some time from three-D images. So for the purpose of diagnosing the diseased areas of three-D image, medical experts need two-D images. This paper addresses the overhaul of three dimensional models from two-D images. In the initial step the image is segmented using level set method. Later segmented image is extracted and registered for overhaul of three dimensional images using metamorphosis and fabric growing methods. The practical result shows the implementation of the suggested method.

Steganography in High Intensity Pixel of a Thermal Image using Barcode Encoder Technique

In this research, the high intensity pixels are the region of interest is selected based on the difference in the intensity level of the colors in an image. The image holding the information is hidden in a cover image which is pattern locked in which a random number is allocated for each of the pixel. The pattern is drawn and the nodes that connect each pixel were selected and marked as the area of interest, which has high intensity as said. These numbers extracted and converted to barcode that are saved as normal image. In this proposed research the information that is hidden in the normal image is fused with the thermal image, which comprises of high intensity colors. The image fusion technique which is proposed in this research is much interesting that ensures more security as it do not reveals any clue about the existence of the information to interpreters. This research, analyzes the Barcode Encoder technique in steganography only on colors with high intensity, and identifies those areas where this technique can be applied, so that the human race could be benefited abundantly.