Search

We propose a straightforward yet highly effective few-shot fine-tuning strategy for adapting the Segment Anything (SAM) to anatomical segmentation tasks in medical images. Our novel approach revolves around reformulating the mask decoder within SAM, leveraging few-shot embeddings derived from a limited set of labeled images (few-shot collection) as prompts for querying anatomical objects captured in image embeddings. This innovative reformulation greatly reduces the need for time-consuming online user interactions for labeling volumetric images, such as exhaustively marking points and bounding boxes to provide prompts slice by slice. With our method, users can manually segment a few 2D slices offline, and the embeddings of these annotated image regions serve as effective prompts for online segmentation tasks. Our method prioritizes the efficiency of the fine-tuning process by exclusively training the mask decoder through caching mechanisms while keeping the image encoder frozen. Importantly, this approach is not limited to volumetric medical images, but can generically be applied to any 2D/3D segmentation task. To thoroughly evaluate our method, we conducted extensive validation on four datasets, covering six anatomical segmentation tasks across two modalities. Furthermore, we conducted a comparative analysis of different prompting options within SAM and the fully-supervised nnU-Net. The results demonstrate the superior performance of our method compared to SAM employing only point prompts (approximately 50% improvement in IoU) and performs on-par with fully supervised methods whilst reducing the requirement of labeled data by at least an order of magnitude., Comment: 9 pages, Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. 2024

Noise plagues many numerical datasets, where the recorded values in the data may fail to match the true underlying values due to reasons including: erroneous sensors, data entry/processing mistakes, or imperfect human estimates. We consider general regression settings with covariates and a potentially corrupted response whose observed values may contain errors. By accounting for various uncertainties, we introduced veracity scores that distinguish between genuine errors and natural data fluctuations, conditioned on the available covariate information in the dataset. We propose a simple yet efficient filtering procedure for eliminating potential errors, and establish theoretical guarantees for our method. We also contribute a new error detection benchmark involving 5 regression datasets with real-world numerical errors (for which the true values are also known). In this benchmark and additional simulation studies, our method identifies incorrect values with better precision/recall than other approaches.

Recent progress in deep generative models has improved the quality of neural vocoders in speech domain. However, generating a high-quality singing voice remains challenging due to a wider variety of musical expressions in pitch, loudness, and pronunciations. In this work, we propose a hierarchical diffusion model for singing voice neural vocoders. The proposed method consists of multiple diffusion models operating in different sampling rates; the model at the lowest sampling rate focuses on generating accurate low-frequency components such as pitch, and other models progressively generate the waveform at higher sampling rates on the basis of the data at the lower sampling rate and acoustic features. Experimental results show that the proposed method produces high-quality singing voices for multiple singers, outperforming state-of-the-art neural vocoders with a similar range of computational costs.

The recent adaption of virtualization technologies in the next generation mobile network enables 5G base station to be segregated into a Radio Unit (RU), a Distributed Unit (DU), and a Central Unit (CU) to support Cloud based Radio Access Networks (C-RAN). RU and DU are connected through a fronthaul link. In contrast, CU and DU are connected through a midhaul link. Although virtualization of CU gives benefits of centralization to the operators, there are other issues to be solved such as optimization of midhaul bandwidth and computing resources at edge cloud and central cloud where the DUs and CUs are deployed, respectively. In this paper, we propose a dynamic functional split selection for the DUs in 5G C-RAN by adopting to traffic heterogeneity where the midhaul bandwidth is limited. We propose an optimization problem that maximizes the centralization of the C-RAN system by operating more number of DUs on split Option-7 by changing the channel bandwidth of the DUs. The dynamic selection of split options among each CU-DU pair gives 90% centralization over the static functional split for a given midhaul bandwidth.

Diwali is among the most important Indian festivals, and elaborate firework displays mark the evening's festivities. This study assesses the impact of Diwali on the concentration, composition, and sources of ambient PM2.5. We observed the total PM2.5 concentrations to rise to 16 times the pre-firework levels, while each of the elemental, organic, and black carbon fractions of ambient PM2.5 increased by a factor of 46.1, 3.7, and 5.6, respectively. The concentration of species like K, Al, Sr, Ba, S, and Bi displayed distinct peaks during the firework event and were identified as tracers. The average concentrations of potential carcinogens, like As, exceeded US EPA screening levels for industrial air by a factor of ~9.6, while peak levels reached up to 16.1 times the screening levels. The source apportionment study, undertaken using positive matrix factorization, revealed the fireworks to account for 95% of the total elemental PM2.5 during Diwali. The resolved primary organic emissions, too, were enhanced by a factor of 8 during Diwali. Delhi has encountered serious haze events following Diwali in recent years; this study highlights that biomass burning emissions rather than the fireworks drive the poor air quality in the days following Diwali., Comment: The manuscript is accepted for publication in Atmospheric Pollution Research. Present Status: Accepted

The Government of India (GOI) announced a nationwide lockdown starting 25th March 2020 to contain the spread of COVID-19, leading to an unprecedented decline in anthropogenic activities and in turn improvements in ambient air quality. This is the first study to focus on highly time-resolved chemical speciation and source apportionment of PM$_{2.5}$ to assess the impact of the lockdown and subsequent relaxations on the sources of ambient PM$_{2.5}$ in Delhi, India. The elemental, organic, and black carbon fractions of PM$_{2.5}$ were measured at the IIT Delhi campus from February 2020 to May 2020. We report source apportionment results using positive matrix factorization (PMF) of organic and elemental fractions of PM$_{2.5}$ during the different phases of the lockdown. The resolved sources such as vehicular emissions, domestic coal combustion, and semi-volatile oxygenated organic aerosol (SVOOA) were found to decrease by 96%, 95%, and 86%, respectively, during lockdown phase-1 as compared to pre-lockdown. An unforeseen rise in O$_3$ concentrations with declining NO$_x$ levels was observed, similar to other parts of the globe, leading to the low-volatility oxygenated organic aerosols (LVOOA) increasing to almost double the pre-lockdown concentrations during the last phase of the lockdown. The effect of the lockdown was found to be less pronounced on other resolved sources like secondary chloride, power plants, dust-related, hydrocarbon-like organic aerosols (HOA), and biomass burning related emissions, which were also swayed by the changing meteorological conditions during the four lockdown phases. The results presented in this study provide a basis for future emission control strategies, quantifying the extent to which constraining certain anthropogenic activities can ameliorate the ambient air., Comment: The article has been accepted for publication in Environment International ISSN: 0160-4120. Please note that the article below has not yet been subjected to formatting and copyediting in line with journal requirements, which may result in some changes prior to publication

With Full Duplex (FD), wireless terminal is capable of transmitting and receiving data simultaneously in the same frequency resources, however, it introduces self interference and co-channel interference. Even though various signal processing techniques are emerged to cancel the self interference, the bottleneck for FD performance in cellular systems is the co-channel interference from the other uplink and downlink signals. In this work we have studied both the uplink and downlink performances of a FD cellular network, where users employ fractional power control in uplink. We use Matern Cluster Process to model the network, which provides a tractable and realistic model to characterize the user-base station distances which are needed for uplink power control. Based on the obtained coverage probabilities, rates and their robust approximations, we show that while FD improves downlink performance, it severely hurts the uplink performance. Also, we provide a trade-off between uplink and downlink performances. Our study suggests dense deployment of low power base stations can improve the performance of FD system., Comment: no longer valid