This paper proposes an explicit way to optimize the super-resolution network for generating visually pleasing images. The previous approaches use several loss functions which is hard to interpret and has the implicit relationships to improve the perceptual score. We show how to exploit the machine learning based model which is directly trained to provide the perceptual score on generated images. It is believed that these models can be used to optimizes the super-resolution network which is easier to interpret. We further analyze the characteristic of the existing loss and our proposed explicit perceptual loss for better interpretation. The experimental results show the explicit approach has a higher perceptual score than other approaches. Finally, we demonstrate the relation of explicit perceptual loss and visually pleasing images using subjective evaluation.
Deep learning models, which are increasingly being used in the field of medical image analysis, come with a major security risk, namely, their vulnerability to adversarial examples. Adversarial examples are carefully crafted samples that force machine learning models to make mistakes during testing time. These malicious samples have been shown to be highly effective in misguiding classification tasks. However, research on the influence of adversarial examples on segmentation is significantly lacking. Given that a large portion of medical imaging problems are effectively segmentation problems, we analyze the impact of adversarial examples on deep learning-based image segmentation models. Specifically, we expose the vulnerability of these models to adversarial examples by proposing the Adaptive Segmentation Mask Attack (ASMA). This novel algorithm makes it possible to craft targeted adversarial examples that come with (1) high intersection-over-union rates between the target adversarial mask and the prediction and (2) with perturbation that is, for the most part, invisible to the bare eye. We lay out experimental and visual evidence by showing results obtained for the ISIC skin lesion segmentation challenge and the problem of glaucoma optic disc segmentation. An implementation of this algorithm and additional examples can be found at https://github.com/utkuozbulak/adaptive-segmentation-mask-attack.
High dimensional data can contain multiple scales of variance. Analysis tools that preferentially operate at one scale can be ineffective at capturing all the information present in this cross-scale complexity. We propose a multiscale joint characterization approach designed to exploit synergies between global and local approaches to dimensionality reduction. We illustrate this approach using Principal Components Analysis (PCA) to characterize global variance structure and t-stochastic neighbor embedding (t-sne) to characterize local variance structure. Using both synthetic images and real-world imaging spectroscopy data, we show that joint characterization is capable of detecting and isolating signals which are not evident from either PCA or t-sne alone. Broadly, t-sne is effective at rendering a randomly oriented low-dimensional map of local clusters, and PCA renders this map interpretable by providing global, physically meaningful structure. This approach is illustrated using imaging spectroscopy data, and may prove particularly useful for other geospatial data given robust local variance structure due to spatial autocorrelation and physical interpretability of global variance structure due to spectral properties of Earth surface materials. However, the fundamental premise could easily be extended to other high dimensional datasets, including image time series and non-image data.
Thoracic disease detection from chest radiographs using deep learning methods has been an active area of research in the last decade. Most previous methods attempt to focus on the diseased organs of the image by identifying spatial regions responsible for significant contributions to the model's prediction. In contrast, expert radiologists first locate the prominent anatomical structures before determining if those regions are anomalous. Therefore, integrating anatomical knowledge within deep learning models could bring substantial improvement in automatic disease classification. This work proposes an anatomy-aware attention-based architecture named Anatomy X-Net, that prioritizes the spatial features guided by the pre-identified anatomy regions. We leverage a semi-supervised learning method using the JSRT dataset containing organ-level annotation to obtain the anatomical segmentation masks (for lungs and heart) for the NIH and CheXpert datasets. The proposed Anatomy X-Net uses the pre-trained DenseNet-121 as the backbone network with two corresponding structured modules, the Anatomy Aware Attention (AAA) and Probabilistic Weighted Average Pooling (PWAP), in a cohesive framework for anatomical attention learning. Our proposed method sets new state-of-the-art performance on the official NIH test set with an AUC score of 0.8439, proving the efficacy of utilizing the anatomy segmentation knowledge to improve the thoracic disease classification. Furthermore, the Anatomy X-Net yields an averaged AUC of 0.9020 on the Stanford CheXpert dataset, improving on existing methods that demonstrate the generalizability of the proposed framework.
The field of neural image compression has witnessed exciting progress as recently proposed architectures already surpass the established transform coding based approaches. While, so far, research has mainly focused on architecture and model improvements, in this work we explore content adaptive optimization. To this end, we introduce an iterative procedure which adapts the latent representation to the specific content we wish to compress while keeping the parameters of the network and the predictive model fixed. Our experiments show that this allows for an overall increase in rate-distortion performance, independently of the specific architecture used. Furthermore, we also evaluate this strategy in the context of adapting a pretrained network to other content that is different in visual appearance or resolution. Here, our experiments show that our adaptation strategy can largely close the gap as compared to models specifically trained for the given content while having the benefit that no additional data in the form of model parameter updates has to be transmitted.
Self-supervised learning on graph-structured data has drawn recent interest for learning generalizable, transferable and robust representations from unlabeled graphs. Among many, graph contrastive learning (GraphCL) has emerged with promising representation learning performance. Unfortunately, unlike its counterpart on image data, the effectiveness of GraphCL hinges on ad-hoc data augmentations, which have to be manually picked per dataset, by either rules of thumb or trial-and-errors, owing to the diverse nature of graph data. That significantly limits the more general applicability of GraphCL. Aiming to fill in this crucial gap, this paper proposes a unified bi-level optimization framework to automatically, adaptively and dynamically select data augmentations when performing GraphCL on specific graph data. The general framework, dubbed JOint Augmentation Optimization (JOAO), is instantiated as min-max optimization. The selections of augmentations made by JOAO are shown to be in general aligned with previous "best practices" observed from handcrafted tuning: yet now being automated, more flexible and versatile. Moreover, we propose a new augmentation-aware projection head mechanism, which will route output features through different projection heads corresponding to different augmentations chosen at each training step. Extensive experiments demonstrate that JOAO performs on par with or sometimes better than the state-of-the-art competitors including GraphCL, on multiple graph datasets of various scales and types, yet without resorting to any laborious dataset-specific tuning on augmentation selection. We release the code at https://github.com/Shen-Lab/GraphCL_Automated.
Multi-task learning (MTL) aims to improve the generalization of several related tasks by learning them jointly. As a comparison, in addition to the joint training scheme, modern meta-learning allows unseen tasks with limited labels during the test phase, in the hope of fast adaptation over them. Despite the subtle difference between MTL and meta-learning in the problem formulation, both learning paradigms share the same insight that the shared structure between existing training tasks could lead to better generalization and adaptation. In this paper, we take one important step further to understand the close connection between these two learning paradigms, through both theoretical analysis and empirical investigation. Theoretically, we first demonstrate that MTL shares the same optimization formulation with a class of gradient-based meta-learning (GBML) algorithms. We then prove that for over-parameterized neural networks with sufficient depth, the learned predictive functions of MTL and GBML are close. In particular, this result implies that the predictions given by these two models are similar over the same unseen task. Empirically, we corroborate our theoretical findings by showing that, with proper implementation, MTL is competitive against state-of-the-art GBML algorithms on a set of few-shot image classification benchmarks. Since existing GBML algorithms often involve costly second-order bi-level optimization, our first-order MTL method is an order of magnitude faster on large-scale datasets such as mini-ImageNet. We believe this work could help bridge the gap between these two learning paradigms, and provide a computationally efficient alternative to GBML that also supports fast task adaptation.
We present a method to learn the 3D surface of objects directly from a collection of images. Previous work achieved this capability by exploiting additional manual annotation, such as object pose, 3D surface templates, temporal continuity of videos, manually selected landmarks, and foreground/background masks. In contrast, our method does not make use of any such annotation. Rather, it builds a generative model, a convolutional neural network, which, given a noise vector sample, outputs the 3D surface and texture of an object and a background image. These 3 components combined with an additional random viewpoint vector are then fed to a differential renderer to produce a view of the sampled object and background. Our general principle is that if the output of the renderer, the generated image, is realistic, then its input, the generated 3D and texture, should also be realistic. To achieve realism, the generative model is trained adversarially against a discriminator that tries to distinguish between the output of the renderer and real images from the given data set. Moreover, our generative model can be paired with an encoder and trained as an autoencoder, to automatically extract the 3D shape, texture and pose of the object in an image. Our trained generative model and encoder show promising results both on real and synthetic data, which demonstrate for the first time that fully unsupervised 3D learning from image collections is possible.
Purpose: To evaluate manual and automatic registration times as well as accuracy with augmented reality during alignment of a holographic 3-dimensional (3D) model onto the real-world environment. Method: 18 participants in various stages of clinical training across two academic centers registered a 3D CT phantom model onto a CT grid using the HoloLens 2 augmented reality headset 3 consecutive times. Registration times and accuracy were compared among different registration methods (hand gesture, Xbox controller, and automatic registration), levels of clinical experience, and consecutive attempts. Registration times were also compared with prior HoloLens 1 data. Results: Mean aggregate manual registration times were 27.7, 24.3, and 72.8 seconds for one-handed gesture, two-handed gesture, and Xbox controller, respectively; mean automatic registration time was 5.3s (ANOVA p<0.0001). No significant difference in registration times was found among attendings, residents and fellows, and medical students (p>0.05). Significant improvements in registration times were detected across consecutive attempts using hand gestures (p<0.01). Compared with previously reported HoloLens 1 experience, hand gesture registration times were 81.7% faster (p<0.05). Registration accuracies were not significantly different across manual registration methods, measuring at 5.9, 9.5, and 8.6 mm with one-handed gesture, two-handed gesture, and Xbox controller, respectively (p>0.05). Conclusions: Manual registration times decreased significantly with updated hand gesture maneuvers on HoloLens 2 versus HoloLens 1, approaching the registration times of automatic registration and outperforming Xbox controller mediated registration. These results will encourage wider clinical integration of HoloLens 2 in procedural medical care.
Recently many first and second order variants of SGD have been proposed to facilitate training of Deep Neural Networks (DNNs). A common limitation of these works stem from the fact that they use the same learning rate across all instances present in the dataset. This setting is widely adopted under the assumption that loss functions for each instance are similar in nature, and hence, a common learning rate can be used. In this work, we relax this assumption and propose an optimization framework which accounts for difference in loss function characteristics across instances. More specifically, our optimizer learns a dynamic learning rate for each instance present in the dataset. Learning a dynamic learning rate for each instance allows our optimization framework to focus on different modes of training data during optimization. When applied to an image classification task, across different CNN architectures, learning dynamic learning rates leads to consistent gains over standard optimizers. When applied to a dataset containing corrupt instances, our framework reduces the learning rates on noisy instances, and improves over the state-of-the-art. Finally, we show that our optimization framework can be used for personalization of a machine learning model towards a known targeted data distribution.