Tadashi Nakano

Attention mechanisms have become an important component in deep learning-based peripheral blood smear analysis. Nevertheless, a major limitation of many existing attention mechanisms is that they generate attention weights from the same feature map they aim to refine. In deep layers, this strategy can produce coarse or spatially misleading attention responses. To address this limitation, this study proposes Cross-level Adaptive Spatial Attention (CASA), an attention mechanism that uses shallow spatial evidence to guide deep feature refinement. CASA first extracts spatial descriptors from shallow feature maps through global and local spatial pooling operations, and then adaptively reduces these descriptors to match the spatial resolution of the target deep feature map before combining with a deep semantic projection to reconstruct spatially reliable attention weights, followed by channel-wise recalibration. By linking shallow spatial details with deep semantic representations, CASA improves the preservation of subtle cellular morphology and reduces interference from noisy background activations. The proposed module is integrated into the early stage of the neck network during feature aggregation in a YOLO-based detector. Experimental results show a strong performance achievement on M5 and MP-IDB datasets that outperform conventional attention modules and several state-of-the-art detectors, with gains of up to 9.8% mAP50. Additional experiments further demonstrate that using the P/4 shallow feature level to guide attention reconstruction provides the most effective balance between spatial detail and semantic reliability. These findings suggest that cross-level spatial attention reconstruction is a promising strategy for robust object detection in noisy peripheral blood smear images.

The increasing use of electric vehicles (EVs) has highlighted the need for sustainable recycling and traceability of essential raw materials. This study introduces a blockchain and AI-integrated framework designed with the possibility to optimize vehicle components lifecycle management with recycling and tracing throughout the supply chain. The system uses Decentralized Identifiers (DIDs) for secure identification of components, enabling transparent tracking from production to recycling. Hyperledger Fabric ensures immutable data an choring across stakeholders. Hyperledger Caliper is used for benchmarking the system, assessing metrics such as transaction speed, latency, and scalability. AI models, including regression and clustering algorithms, are utilized to optimize recycling processes, predict component lifespans, and enhance resource recovery in the system. A tokenized reward mechanism incentivizes eco-friendly practices among stakeholders. The system also shows its environmental benefits, including energy savings from improved recycling performance. The proposed framework effectively supports the circular economy by enhancing resource recovery processes. While its design has the potential to reduce environmental impact, this benefit depends on the reusable, modular system design.

Multicellular systems form diverse morphologies through chemical and physical interactions among cells. In this study, we propose a trainable multicellular phase-field model that incorporates a neural network. The phase-field model represents cell morphology and position as continuous fields and serves as the basis for learning intercellular dynamics. In the proposed model, the neural network takes the local state as input and predicts the subsequent state, adjusting the spatial evolution of the multicellular system to approach a target pattern defined by a loss function. Through simulation experiments, we confirmed that the trained neural network enables cells to aggregate into a cluster and reproduce the spatial pattern of a given image. Moreover, similar results were obtained in a larger simulation domain than that used during training, demonstrating the scalability of the model.

We present a computational model for studying morphogenesis based on pairwise interactions among cells in this article. Extending a previous single-particle model, we represent each cell using two particles connected by an internal link. This two-particle model introduces cell shape and orientation while retaining the simplicity of a force-based framework. The model includes a remote force acting between all cell pairs and a contact force acting only between physically connected pairs. Simulation results showed that the model reproduces a range of multicellular patterns depending on cell density and interaction parameters. This approach provides a minimal and extensible framework for investigating the physical basis of tissue-level organization.