Sharma Meghna, Arora Shaveta
Department of Computer Science and Engineering, The NorthCap University, Gurugram, Haryana, India.
Psychiatry Res Neuroimaging. 2025 Aug;351:112016. doi: 10.1016/j.pscychresns.2025.112016. Epub 2025 Jun 13.
As deep learning continues to advance in medical analysis, the increasing complexity of models, particularly Convolutional Neural Networks (CNNs), presents significant challenges related to interpretability, computational costs, and real-world applicability. These issues are critical in the medical domain, e.g., Attention Deficit Hyperactivity Disorder (ADHD) diagnosis, where model efficiency and interpretability are paramount. This paper proposes a novel parameter-efficient framework based on the Kolmogorov-Arnold Network (KAN) to overcome these challenges. Unlike CNNs, KAN restructures feature transformations, significantly reducing parameter overhead while preserving high classification accuracy. An attention-driven feature selection mechanism dynamically prioritizes the most significant features, minimizing irrelevant features and unnecessary computational load. Recognizing the complex and diverse nature of ADHD- related brain connectivity features, a novel activation function with learnable coefficients is introduced, enabling adaptive transformation based on specific data patterns. To further enhance model generalization, an advanced sliding window-based data augmentation technique is incorporated to meet substantial data requirements for training. Extensive experimentation on the benchmark ADHD-200 dataset demonstrates the model's superiority, achieving an accuracy of 79.25 %, an F1-score of 78. 75 % and a precision of 78.23 %, surpassing many state-of-the-art ADHD studies. Remarkably, these results are achieved using only a few thousand parameters compared to the millions required by many existing approaches, making it valuable for various resource-constrained researchers and organizations. The proposed framework, seamlessly fusing KAN, attention-driven feature selection, adaptive activation, and robust data augmentation, achieves substantial parameter reduction with enhanced performance. This lightweight architecture, combined with superior performance and interpretability, makes the proposed model highly promising for ADHD diagnosis and other complex medical applications.
随着深度学习在医学分析领域不断发展，模型的复杂性日益增加，尤其是卷积神经网络（CNN），这在可解释性、计算成本和实际适用性方面带来了重大挑战。这些问题在医学领域至关重要，例如注意力缺陷多动障碍（ADHD）诊断，其中模型效率和可解释性至关重要。本文提出了一种基于柯尔莫哥洛夫 - 阿诺德网络（KAN）的新型参数高效框架，以克服这些挑战。与CNN不同，KAN重新构建了特征变换，在保持高分类准确率的同时显著减少了参数开销。一种注意力驱动的特征选择机制动态地对最重要的特征进行优先级排序，最大限度地减少无关特征和不必要的计算负载。认识到与ADHD相关的大脑连接特征的复杂多样性，引入了一种具有可学习系数的新型激活函数，能够基于特定数据模式进行自适应变换。为了进一步提高模型的泛化能力，采用了一种先进的基于滑动窗口的数据增强技术，以满足训练所需的大量数据要求。在基准ADHD - 200数据集上进行的广泛实验证明了该模型的优越性，准确率达到79.25%，F1分数为78.75%，精确率为78.23%，超过了许多现有最先进的ADHD研究。值得注意的是，与许多现有方法所需的数百万参数相比，这些结果仅使用了几千个参数就实现了，这使其对各种资源受限的研究人员和组织具有价值。所提出的框架无缝融合了KAN、注意力驱动的特征选择、自适应激活和强大的数据增强，在提高性能的同时实现了显著的参数减少。这种轻量级架构，结合卓越的性能和可解释性，使得所提出的模型在ADHD诊断和其他复杂医学应用中极具前景。
