Articles
Publication Date: 2025
Engineering Applications of Artificial Intelligence (09521976)156
Cellular heterogeneity, even among genetically identical cells, results in variations in their properties and behaviors, making single-cell analysis crucial for obtaining detailed insights. However, isolating single cells from a cell population poses major challenges, as conventional laboratory techniques often risk cell damage and involve complex procedures. Droplet microfluidics has emerged as a promising approach for encapsulating cells, particularly single cells, into individual droplets without causing harm. Despite this, factors like cell sedimentation and aggregation can reduce encapsulation efficiency and lead to deviations from the expected Poisson distribution. To address these challenges, leveraging artificial intelligence and deep learning to monitor, detect, and regulate encapsulation conditions in real-time is critical for enhancing system performance. However, deep learning models require substantial training data, and issues like microfluidic channel clogging and the scarcity of certain cell types often limit data availability. To overcome this limitation, researchers are turning to synthetic data generation to supplement training datasets and address data scarcity challenges effectively. This study emphasizes the potential of integrating synthetic data with cutting-edge deep learning techniques to enhance the accuracy and efficiency of single-cell analysis within droplet microfluidic systems. A diverse dataset integrating synthetic and real images was used to train the YOLOv8s model for automated detection and classification of microfluidic droplets, enhancing accuracy and system performance. The model trained on a combination of real and synthetic data outperformed the one trained using conventional data augmentation methods, achieving an mAP0.5 of 98% due to the increased diversity of training images. It also demonstrated faster and more stable training. Additionally, the YOLOv8 network, with a detection rate of approximately 2338 droplets per second, significantly improved processing speed compared to earlier YOLO versions. © 2025
Publication Date: 2025
Journal of Combinatorial Optimization (13826905)49(1)
In this work, we consider a combinatorial optimization problem with direct applications in blockchain mining, namely finding the most lucrative blocks for Bitcoin miners, and propose optimal algorithmic solutions. Our experiments show that our algorithms increase the miners’ revenues by more than a million dollars per month. Modern blockchains reward their miners in two ways: (i) a base reward for each block that is mined, and (ii) the transaction fees of those transactions that are included in the mined block. The base reward is fixed by the respective blockchain’s protocol and is not under the miner’s control. Hence, for a miner who wishes to maximize earnings, the fundamental problem is to form a valid block with maximal total transaction fees and then try to mine it. Moreover, in many protocols, including Bitcoin itself, the base reward halves at predetermined intervals, hence increasing the importance of maximizing transaction fees and mining an optimal block. This problem is further complicated by the fact that transactions can be prerequisites of each other or have conflicts (in case of double-spending). In this work, we consider the problem of forming an optimal block, i.e. a valid block with maximal total transaction fees, given a set of unmined transactions. On the theoretical side, we first formally model our problem as an extension of Knapsack and then show that, unlike classical Knapsack, our problem is strongly NP-hard. We also show a hardness-of-approximation result. As such, there is no hope in solving it efficiently for general instances. However, we observe that its real-world instances are quite sparse, i.e. the transactions have very few dependencies and conflicts. Using this fact, and exploiting three well-known graph sparsity parameters, namely treedepth, treewidth and pathwidth, we present exact linear-time parameterized algorithms that are applicable to the real-world instances and obtain optimal results. On the practical side, we provide an extensive experimental evaluation demonstrating that our approach vastly outperforms the current Bitcoin miners in practice, obtaining a significant per-block average increase of 11.34 percent in transaction fee revenues which amounts to almost one million dollars per month. © The Author(s) 2024.
