摘要
准确检测胃肠道(GI)疾病对有效的医疗干预至关重要。现有的方法往往缺乏准确性和效率，强调需要更先进的方法。医学图像数据的复杂性和多样性，如在胃肠道疾病中发现的数据，可能对单一模型来全面表示所有基本特征提出挑战。在这种情况下，集成学习方法变得非常重要。本文提出一种新颖的集成学习方法用于GI疾病预测。我们利用DenseNet169、InceptionV3和MobileNet这三种传输学习模型的强大功能作为基础学习者，并利用其他层有效地学习特定于数据的fc功能。我们实现了加权平均集成策略来合并来自单个基础模型的预测，并使用布谷鸟搜索(CS)和Levy Fight算法来调整权重。与单个模型相比，这种方法导致了更准确的预测，利用基础学习者的不同优势来提高胃肠道疾病预测的性能。这项研究是引入基于元启发式的优化模型以检测胃肠道疾病的先驱。我们使用由6000张图像组成的可公开访问的内窥镜图像数据集来评估所提出的模型。结果显示了异常的预测精度，其中集成达到了99.75%的突出精度。通过Grad-CAM分析，我们对个别基础模型的决策过程获得了宝贵的见解，使我们能够确定优势和改进的领域。我们提出的集成模型超越了传统的权重分配方法和现有的最先进的方法，展示了它在胃肠道疾病预测中的优越性。我们的方法在医学图像分析方面具有变革性的潜力，有望增强胃肠病学的患者护理和诊断准确性。
Abstract
Accurate detection of gastrointestinal (GI) diseases is critical for efective medical inter- vention. Existing methods often lack accuracy and efciency, emphasizing the need for more advanced approaches. The complexity and diversity of medical image data, such as those found in GI diseases, can pose challenges for a single model to comprehensively represent all essential features. In such scenarios, an ensemble learning approach becomes important. In this paper, we propose an innovative ensemble learning approach for GI dis- ease prediction. We leverage the power of three transfer learning models, DenseNet169, InceptionV3, and MobileNet, as base learners along with additional layers to efectively learn data-specifc features. We implement a weighted averaging ensemble strategy to merge predictions from individual base models and fne-tune the weights using the cuckoo search (CS) with levy fight algorithm. This approach results in more accurate predictions compared to individual models, leveraging the diverse strengths of the base learners for enhanced performance in GI disease prediction. This study is notably the pioneer in intro- ducing a metaheuristics-based optimized model for the detection of GI diseases. We assess the presented model using a publicly accessible endoscopic image dataset that consists of 6,000 images. The results demonstrate exceptional predictive accuracy, with the ensem- ble achieving an outstanding accuracy of 99.75%. Through Grad-CAM analysis, we gain valuable insights into the decision-making process of the individual base models, enabling us to identify areas of strength and improvement. Our proposed ensemble model outper- forms traditional weight assignment methods and existing state-of-the-art methods, show- casing its superiority in GI disease prediction. Our approach has transformative potential in medical image analysis, promising enhanced patient care and diagnostic accuracy in gastroenterology.
NETL
NSTL