China 
Africa 
Explore chapters and articles related to this topic
Introduction to Deep Learning
Published in Lia Morra, Silvia Delsanto, Loredana Correale, Artificial Intelligence in Medical Imaging, 2019
Lia Morra, Silvia Delsanto, Loredana Correale
Another flourishing line of research leverages patient-level annotations to train lesion-level classifiers. In this setting, we know that an abnormality is present in an image, but its precise location is unknown, yet we wish to predict it. In machine learning, this setting is known as multiple-instance learning (MIL), in which the learner receives a set of labeled bags, each containing many (unlabelled) instances. In the simplest case of binary classification, a bag is labeled positive if it contains at least one positive instance. MIL is a natural learning scenario for medical image analysis because labels are often not available at the desired granularity. And has been applied to several application domains including histopathology analysis, diabetic retinopathy, and lesion detection in lung, abdomen and breast images [96].
A scaling up approach: a research agenda for medical imaging analysis with applications in deep learning
Published in Journal of Experimental & Theoretical Artificial Intelligence, 2023
Yaw Afriyie, Benjamin A. Weyori, Alex A. Opoku
While large amounts of unlabelled medical images are more publicly available, the absence of labelled data is often a challenge when machine learning methods are applied (Litjens et al., 2017; Weese & Lorenz, 2016), (Ørting et al., 2019). There are several reasons for this, one being the difficulty in gathering the labels. The process of manually categorising images is costly and/or time-consuming. The use of labelled data for clinical practice may not be necessary, limiting its application to research investigations only. Even when labels are applied to data, it is rarely made accessible to other researchers (Ørting et al., 2019). In the absence of labelled data, techniques that contribute additional data or labels are encouraged to supplement typical supervised learning. There are a wide variety of terms used to describe these approaches, including semi-supervised learning, multiple instance learning, and transfer learning. These research papers appear to be aware of relevant literature, and surveys appear, such as this one (Quellec et al., 2012). This study answers the question ‘Can image identification and classification with deep learning algorithms improve medical imaging analysis? ’ since the prior work done are unable to identify commonalities between deep learning tools and their designs. Despite their similar objectives, there does not appear to be any interaction between the scenarios.
Visual Tracking Using Kernelized Correlation Filter with Conditional Switching to Median Flow Tracker
Published in IETE Journal of Research, 2020
C. S. Asha, A. V. Narasimhadhan
In literature, two types of trackers have been proposed namely, generative and discriminative techniques. In generative techniques, tracker uses the information of an object to search for the most similar template in every frame. They include template matching [4], colour histogram matching [5], subspace techniques such as incremental principal component analysis [6], particle filter [7] and sparse representation using dictionary [8], optical flow based method [9], spatial histogram matching [10,11] to track an object. In contrary, discriminative trackers learn the features of an object and background from online data. In this, the problem is treated as a binary classification task to locate the object in every frame. Such approaches have been proposed by training the edge feature with support vector machine (SVM) classifier [12], Haar features with online AdaBoost classifier [13], multiple instance learning using Haar features of an object and background [14], 2-bit pattern with the random fern classifier as a detector and integrated with optical flow tracker [15], Haar features of an object with Bayes classifier [16], Haar features, raw features, and histogram features trained with structured SVM classifier for an efficient tracking [17]. Recently, deep learning techniques have been employed in tracking [18,19], but real-time implementation is still a challenge. In addition, they demand huge data and large computational power.
An intelligent unsupervised anomaly detection in videos using inception capsule auto encoder
Published in The Imaging Science Journal, 2023
Harshadkumar S. Modi, Dhaval A. Parikh
Sarker et al. [31] presented a semi-supervised anomaly detection framework in wild video sequences. Here, a temporal convolutional 3D neural network (T-C3D) method and Multiple Instance Learning (MIL) classifiers are introduced to generate automatic anomaly detection. This work develops a ranking loss function, which enhances the interval among the score of classification anomalous and normal input videos. This ranking loss function diminishes the false negatives and gives the system higher results. Nevertheless, the established approach faces issues maintaining frames and videos with reduced resolution, fast motions, and the worst illumination. Hence, it affects the efficacy of the developed model.