The medical imaging literature has witnessed remarkable progress in high-performing segmentation models based on convolutional neural networks. Despite the new performance highs, the recent advanced segmentation models still require large, representative, and high quality annotated datasets. However, rarely do we have a perfect training dataset, particularly in the field of medical imaging, where data and annotations are both expensive to acquire. Recently, a large body of research has studied the problem of medical image segmentation with imperfect datasets, tackling two major dataset limitations: scarce annotations where only limited annotated data is available for training, and weak annotations where the training data has only sparse annotations, noisy annotations, or image-level annotations. In this article, we provide a detailed review of the solutions above, summarizing both the technical novelties and empirical results. We further compare the benefits and requirements of the surveyed methodologies and provide our recommended solutions to the problems of scarce and weak annotations. We hope this review increases the community awareness of the techniques to handle imperfect datasets.

Working memory (WM) mechanisms for verbal, spatial, and object information have been extensively examined, yet those for kinetic information are less known. The current study explored the WM capacity and architecture of kinetic information by examining the maintenance of biological motion (BM) stimuli in WM. Human BM is the most salient and biologically significant kinetic information encountered in everyday life. We isolated motion signals of human BM from non-BM sources by using point-light displays as to-be-memorized BM. During a change detection task, we found that, at most, 3 to 4 BM stimuli could be retained in WM (Experiment 1). Next, we found that extra colors, spatial locations, or shapes remembered concurrently with BM stimuli (Experiments 2, 3, and 4, respectively), did not affect BM memory considerably. However, BM memory was affected by a concurrent memory task of non-BM movements (Experiment 5). These results support the hypothesis that an independent storage buffer of WM exists for kinetic information, which can hold up to 3 to 4 motion units.

Working memory mechanisms for binding have been examined extensively in the last decade, yet few studies have explored bindings relating to human biological motion (BM). Human BM is the most salient and biologically significant kinetic information encountered in everyday life and is stored independently from other visual features (e.g., colors). The current study explored 3 critical issues of BM-related binding in working memory: (a) how many BM binding units can be retained in working memory, (b) whether involuntarily object-based binding occurs during BM binding, and (c) whether the maintenance of BM bindings in working memory requires attention above and beyond that needed to maintain the constituent dimensions. We isolated motion signals of human BM from non-BM sources by using point-light displays as to-be-memorized BM and presented the participants colored BM in a change detection task. We found that working memory capacity for BM-color bindings is rather low; only 1 or 2 BM-color bindings could be retained in working memory regardless of the presentation manners (Experiments 1-3). Furthermore, no object-based encoding took place for colored BM stimuli regardless of the processed dimensions (Experiments 4 and 5). Central executive attention contributes to the maintenance of BM-color bindings, yet maintaining BM bindings in working memory did not require more central attention than did maintaining the constituent dimensions in working memory (Experiment 6). Overall, these results suggest that keeping BM bindings in working memory is a fairly resource-demanding process, yet central executive attention does not play a special role in this cross-module binding.

Every day, people perceive other people performing interactive actions. Retaining these actions of human agents in working memory (WM) plays a pivotal role in a normal social life. However, whether the semantic knowledge embedded in the interactive actions has a pervasive impact on the storage of the actions in WM remains unknown. In the current study, we investigated two opposing hypotheses: (a) that WM stores the interactions individually (the individual-storage hypothesis) and (b) that WM stores the interactions as chunks (the chunk-storage hypothesis). We required participants to memorize a set of individual actions while ignoring the underlying social interactions. We found that although the social-interaction aspect was task irrelevant, the interactive actions were stored in WM as chunks that were not affected by memory load (Experiments 1 and 2); however, inverting the human actions vertically abolished this chunking effect (Experiment 3). These results suggest that WM automatically and efficiently used semantic knowledge about interactive actions to store them and support the chunk-storage hypothesis.

No abstract

No abstract

We investigate the effectiveness of a simple solution to the common problem of deep learning in medical image analysis with limited quantities of labeled training data. The underlying idea is to assign artificial labels to abundantly available unlabeled medical images and, through a process known as surrogate supervision, pre-train a deep neural network model for the target medical image analysis task lacking sufficient labeled training data. In particular, we employ 3 surrogate supervision schemes, namely rotation, reconstruction, and colorization, in 4 different medical imaging applications representing classification and segmentation for both 2D and 3D medical images. 3 key findings emerge from our research: 1) pre-training with surrogate supervision is effective for small training sets; 2) deep models trained from initial weights pre-trained through surrogate supervision outperform the same models when trained from scratch, suggesting that pretraining with surrogate supervision should be considered prior to training any deep 3D models; 3) pre-training models in the medical domain with surrogate supervision is more effective than transfer learning from an unrelated domain (e.g., natural images), indicating the practical value of abundant unlabeled medical image data.

BackgroundThis study explored whether the high-resolution representations created by visual working memory (VWM) are constructed in a coarse-to-fine or all-or-none manner. The coarse-to-fine hypothesis suggests that coarse information precedes detailed information in entering VWM and that its resolution increases along with the processing time of the memory array, whereas the all-or-none hypothesis claims that either both enter into VWM simultaneously, or neither does.Methodology/Principal FindingsWe tested the two hypotheses by asking participants to remember two or four complex objects. An ERP component, contralateral delay activity (CDA), was used as the neural marker. CDA is higher for four objects than for two objects when coarse information is primarily extracted; yet, this CDA difference vanishes when detailed information is encoded. Experiment 1 manipulated the comparison difficulty of the task under a 500-ms exposure time to determine a condition in which the detailed information was maintained. No CDA difference was found between two and four objects, even in an easy-comparison condition. Thus, Experiment 2 manipulated the memory array’s exposure time under the easy-comparison condition and found a significant CDA difference at 100 ms while replicating Experiment 1′s results at 500 ms. In Experiment 3, the 500-ms memory array was blurred to block the detailed information; this manipulation reestablished a significant CDA difference.Conclusions/SignificanceThese findings suggest that the creation of high-resolution representations in VWM is a coarse-to-fine process.