Intelligence-based medicine最新文献_第4页

Regulating intelligence: a systematic analysis of safety, ethics, and equity in artificial intelligence driven healthcare 规范智能：对人工智能驱动的医疗保健中的安全、伦理和公平进行系统分析

Intelligence-based medicine

Pub Date : 2025-01-01 DOI: 10.1016/j.ibmed.2025.100320

Muayyad Ahmad

Objective

The growing use of artificial intelligence (AI) in healthcare demands robust regulatory frameworks to ensure safety, ethics, and legal compliance, particularly regarding algorithmic transparency, data privacy, and bias. This systematic review analyzes AI regulations and grey literature (2019–2024) from the Food and Drug Administration (FDA), the World Health Organization (WHO), the Organization for Economic Co-operation and Development (OECD), and the International Organization for Standardization/International Electrotechnical Commission (ISO/IEC).

Methods

This systematic review, following Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines, analysed 26 studies (Peer-reviewed studies: n = 17 & Grey literature: n = 9) on AI regulatory frameworks in healthcare between 2019 and 2024. A systematic literature search and rigorous inclusion criteria ensured relevance, with data consolidated into safety, effectiveness, and ethical themes. The analysis integrates a low-middle income countries (LMIC) perspective via WHO policy and a handful of studies, but the main academic body is drawn primarily from high-income contexts.

Bias risk was assessed systematically.

Results

The reviewed studies highlight the critical need for AI regulatory frameworks in healthcare, focusing on patient safety, ethics, and trust. Key findings stress the necessity for transparent, equitable integration and clear guidelines addressing bias, legal issues, and validation. Grey literature consistently emphasizes risk-based safety models and principles like transparency and human oversight. However, a significant gap remains in translating equity commitments into enforceable standards for bias mitigation, underscoring a critical need for future regulatory action.

Conclusion

This review identifies critical gaps in AI regulatory frameworks, particularly in equity, real-world validation, and liability, and proposes actionable, interdisciplinary strategies to ensure AI's safe, ethical, and equitable integration into healthcare.

人工智能（AI）在医疗保健领域的日益广泛应用需要强有力的监管框架，以确保安全、道德和法律合规性，特别是在算法透明度、数据隐私和偏见方面。本系统综述分析了美国食品药品监督管理局（FDA）、世界卫生组织（WHO）、经济合作与发展组织（OECD）和国际标准化组织/国际电工委员会（ISO/IEC）的人工智能法规和灰色文献（2019-2024）。方法：本系统评价遵循系统评价和荟萃分析（PRISMA）指南的首选报告项目，分析了26项关于2019年至2024年医疗保健领域人工智能监管框架的研究（同行评审研究：n = 17；灰色文献：n = 9）。系统的文献检索和严格的纳入标准确保了相关性，并将数据整合到安全性、有效性和伦理主题中。该分析通过世卫组织政策和少数研究整合了中低收入国家的视角，但主要学术机构主要来自高收入背景。系统评估偏倚风险。结果：回顾的研究强调了在医疗保健领域建立人工智能监管框架的迫切需要，重点是患者安全、道德和信任。主要结论强调了透明、公平整合的必要性，以及解决偏见、法律问题和有效性的明确指导方针。灰色文献一贯强调基于风险的安全模型和原则，如透明度和人为监督。然而，在将股权承诺转化为可执行的减轻偏见标准方面仍存在重大差距，这突出表明迫切需要今后采取监管行动。本综述确定了人工智能监管框架的关键差距，特别是在公平、现实验证和责任方面，并提出了可操作的跨学科战略，以确保人工智能安全、道德和公平地融入医疗保健。

{"title":"Regulating intelligence: a systematic analysis of safety, ethics, and equity in artificial intelligence driven healthcare","authors":"Muayyad Ahmad","doi":"10.1016/j.ibmed.2025.100320","DOIUrl":"10.1016/j.ibmed.2025.100320","url":null,"abstract":"<div><h3>Objective</h3><div>The growing use of artificial intelligence (AI) in healthcare demands robust regulatory frameworks to ensure safety, ethics, and legal compliance, particularly regarding algorithmic transparency, data privacy, and bias. This systematic review analyzes AI regulations and grey literature (2019–2024) from the Food and Drug Administration (FDA), the World Health Organization (WHO), the Organization for Economic Co-operation and Development (OECD), and the International Organization for Standardization/International Electrotechnical Commission (ISO/IEC).</div></div><div><h3>Methods</h3><div>This systematic review, following Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines, analysed 26 studies (Peer-reviewed studies: n = 17 & Grey literature: n = 9) on AI regulatory frameworks in healthcare between 2019 and 2024. A systematic literature search and rigorous inclusion criteria ensured relevance, with data consolidated into safety, effectiveness, and ethical themes. The analysis integrates a low-middle income countries (LMIC) perspective via WHO policy and a handful of studies, but the main academic body is drawn primarily from high-income contexts.</div><div>Bias risk was assessed systematically.</div></div><div><h3>Results</h3><div>The reviewed studies highlight the critical need for AI regulatory frameworks in healthcare, focusing on patient safety, ethics, and trust. Key findings stress the necessity for transparent, equitable integration and clear guidelines addressing bias, legal issues, and validation. Grey literature consistently emphasizes risk-based safety models and principles like transparency and human oversight. However, a significant gap remains in translating equity commitments into enforceable standards for bias mitigation, underscoring a critical need for future regulatory action.</div></div><div><h3>Conclusion</h3><div>This review identifies critical gaps in AI regulatory frameworks, particularly in equity, real-world validation, and liability, and proposes actionable, interdisciplinary strategies to ensure AI's safe, ethical, and equitable integration into healthcare.</div></div>","PeriodicalId":73399,"journal":{"name":"Intelligence-based medicine","volume":"12 ","pages":"Article 100320"},"PeriodicalIF":0.0,"publicationDate":"2025-01-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145618006","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

3D brain tumor segmentation in MRI images using hierarchical adaptive pruning of non-tumor regions 基于非肿瘤区域分层自适应剪枝的MRI图像三维脑肿瘤分割

Intelligence-based medicine

Pub Date : 2025-01-01 DOI: 10.1016/j.ibmed.2025.100303

Ali Mehrabi, Nasser Mehrshad

Background

The detection of brain tumors in MRI images has significantly improved with the advent of deep learning methods. However, these approaches often suffer from high complexity, computational cost, and the need for extensive annotated training data, making them less practical for real-time and patient-centered diagnostic systems. To address these challenges, this study introduces a perceptually inspired, algorithmic method that mimics the diagnostic behavior of physicians, offering a lightweight and interpretable alternative for brain tumor segmentation.

Method

We propose a novel adaptive hierarchical pruning algorithm for 3D MRI brain images that iteratively removes low-intensity, non-tumor voxels based on the statistical distribution of intensities. The tumor region is identified through the comparison of the remaining pixel intensity values statistics. The pruning automatically stops when the mean and median of the remaining voxels converge, leaving the candidate tumor region.

Results

The proposed algorithm was evaluated on all patients of the BraTS2019 and BraTS2023 datasets, achieving segmentation accuracies of 99.1 % and 99.13 %, respectively. It demonstrated high sensitivity and specificity compared to several deep learning methods, showing robust performance across diverse patient scans.

Conclusions

This study demonstrates that a simple, perceptually driven segmentation algorithm can match or outperform complex deep learning models, particularly in clinical settings where lightweight, transparent, and efficient tools are essential.

随着深度学习方法的出现，MRI图像中脑肿瘤的检测有了显著的提高。然而，这些方法通常存在复杂性高、计算成本高、需要大量带注释的训练数据等问题，这使得它们在实时和以患者为中心的诊断系统中不太实用。为了解决这些挑战，本研究引入了一种感知启发的算法方法，该方法模仿医生的诊断行为，为脑肿瘤分割提供了一种轻量级且可解释的替代方法。方法提出了一种新的自适应分层剪枝算法，该算法基于强度的统计分布，迭代地去除低强度的非肿瘤体素。通过比较剩余的像素强度值统计来识别肿瘤区域。当剩余体素的均值和中值收敛时，剪枝自动停止，留下候选肿瘤区域。结果该算法在BraTS2019和BraTS2023数据集的所有患者上进行了评估，分割准确率分别达到99.1%和99.13%。与几种深度学习方法相比，它表现出高灵敏度和特异性，在不同的患者扫描中表现出强大的性能。本研究表明，一个简单的、感知驱动的分割算法可以匹配或优于复杂的深度学习模型，特别是在临床环境中，轻量级、透明和高效的工具是必不可少的。

{"title":"3D brain tumor segmentation in MRI images using hierarchical adaptive pruning of non-tumor regions","authors":"Ali Mehrabi, Nasser Mehrshad","doi":"10.1016/j.ibmed.2025.100303","DOIUrl":"10.1016/j.ibmed.2025.100303","url":null,"abstract":"<div><h3>Background</h3><div>The detection of brain tumors in MRI images has significantly improved with the advent of deep learning methods. However, these approaches often suffer from high complexity, computational cost, and the need for extensive annotated training data, making them less practical for real-time and patient-centered diagnostic systems. To address these challenges, this study introduces a perceptually inspired, algorithmic method that mimics the diagnostic behavior of physicians, offering a lightweight and interpretable alternative for brain tumor segmentation.</div></div><div><h3>Method</h3><div>We propose a novel adaptive hierarchical pruning algorithm for 3D MRI brain images that iteratively removes low-intensity, non-tumor voxels based on the statistical distribution of intensities. The tumor region is identified through the comparison of the remaining pixel intensity values statistics. The pruning automatically stops when the mean and median of the remaining voxels converge, leaving the candidate tumor region.</div></div><div><h3>Results</h3><div>The proposed algorithm was evaluated on all patients of the BraTS2019 and BraTS2023 datasets, achieving segmentation accuracies of 99.1 % and 99.13 %, respectively. It demonstrated high sensitivity and specificity compared to several deep learning methods, showing robust performance across diverse patient scans.</div></div><div><h3>Conclusions</h3><div>This study demonstrates that a simple, perceptually driven segmentation algorithm can match or outperform complex deep learning models, particularly in clinical settings where lightweight, transparent, and efficient tools are essential.</div></div>","PeriodicalId":73399,"journal":{"name":"Intelligence-based medicine","volume":"12 ","pages":"Article 100303"},"PeriodicalIF":0.0,"publicationDate":"2025-01-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145319572","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Segmentation of low-resolution MII human oocyte images using data-efficient meta-learning 使用数据高效元学习的低分辨率MII人类卵母细胞图像分割

Intelligence-based medicine

Pub Date : 2025-01-01 DOI: 10.1016/j.ibmed.2025.100327

Mahshid Alizadeh Kiashi, Ashkan Mousazadeh Soustani, Seyed Abolghasem Mirroshandel

Although many infertility problems in humans are treatable today, some of these methods have problems that need to be addressed. One of the ways to solve infertility problems is to use in vitro fertilization. In this method, if a suitable oocyte with a high chance of fertility is not selected, it may lead to multiple births or even infertility. Identifying the most suitable cell with the highest chance of fertility is a very difficult task even for embryologists. Hence, this research focuses on designing a deep learning framework to take on the challenging task of segmenting low-resolution oocyte microscopic images and accurately delineating and distinguishing the boundaries of the three important regions of the oocyte, i.e., zona pellucida (ZP), perivitelline space (PVS), and ooplasm. Then it is calculated with the key indicators in the health of each cell and compared with the true values to evaluate the level of abnormalities and select the most suitable cell for in vitro fertilization. Finally, after testing on 253 images of the test set, in binary segmentation, the average accuracy is 99.17 in ooplasm, 97.14 in PVS, and 94.78 in ZP, and in multi-class segmentation, the average accuracy is 99.30 in ooplasm, 97.36 in PVS, and 94.95 in ZP. These values have been obtained by training on 300 microscopic images of human oocytes, which have been reduced to less than half compared to previous studies.

虽然今天人类的许多不孕症是可以治疗的，但其中一些方法存在需要解决的问题。解决不孕问题的方法之一是使用体外受精。在这种方法中，如果没有选择生育机会高的合适卵母细胞，可能会导致多胎甚至不孕。即使对胚胎学家来说，确定最合适的细胞和最高的生育机会也是一项非常困难的任务。因此，本研究的重点是设计一个深度学习框架，以承担低分辨率卵母细胞显微图像的分割任务，并准确描绘和区分卵母细胞的三个重要区域，即透明带（ZP）、卵泡周围空间（PVS）和卵浆的边界。然后与各细胞健康状况的关键指标进行计算，并与真实值进行比较，评估异常程度，选择最适合体外受精的细胞。最后，对测试集的253张图像进行测试，在二值分割中，卵浆的平均准确率为99.17，PVS的平均准确率为97.14，ZP的平均准确率为94.78；在多类分割中，卵浆的平均准确率为99.30，PVS的平均准确率为97.36，ZP的平均准确率为94.95。这些值是通过对300个人类卵母细胞的显微图像进行训练获得的，与以前的研究相比，这些图像减少到不到一半。

{"title":"Segmentation of low-resolution MII human oocyte images using data-efficient meta-learning","authors":"Mahshid Alizadeh Kiashi, Ashkan Mousazadeh Soustani, Seyed Abolghasem Mirroshandel","doi":"10.1016/j.ibmed.2025.100327","DOIUrl":"10.1016/j.ibmed.2025.100327","url":null,"abstract":"<div><div>Although many infertility problems in humans are treatable today, some of these methods have problems that need to be addressed. One of the ways to solve infertility problems is to use in vitro fertilization. In this method, if a suitable oocyte with a high chance of fertility is not selected, it may lead to multiple births or even infertility. Identifying the most suitable cell with the highest chance of fertility is a very difficult task even for embryologists. Hence, this research focuses on designing a deep learning framework to take on the challenging task of segmenting low-resolution oocyte microscopic images and accurately delineating and distinguishing the boundaries of the three important regions of the oocyte, i.e., zona pellucida (ZP), perivitelline space (PVS), and ooplasm. Then it is calculated with the key indicators in the health of each cell and compared with the true values to evaluate the level of abnormalities and select the most suitable cell for in vitro fertilization. Finally, after testing on 253 images of the test set, in binary segmentation, the average accuracy is 99.17 in ooplasm, 97.14 in PVS, and 94.78 in ZP, and in multi-class segmentation, the average accuracy is 99.30 in ooplasm, 97.36 in PVS, and 94.95 in ZP. These values have been obtained by training on 300 microscopic images of human oocytes, which have been reduced to less than half compared to previous studies.</div></div>","PeriodicalId":73399,"journal":{"name":"Intelligence-based medicine","volume":"12 ","pages":"Article 100327"},"PeriodicalIF":0.0,"publicationDate":"2025-01-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145683784","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Predicting COPD admissions using machine learning and SHAP: An exploratory multi-hospital study in Riyadh, Saudi Arabia 使用机器学习和SHAP预测COPD入院：沙特阿拉伯利雅得的一项探索性多医院研究

Intelligence-based medicine

Pub Date : 2025-01-01 DOI: 10.1016/j.ibmed.2025.100312

Anas Ali Alhur , Jamilu Sani , Mohamed Mustaf Ahmed

Background

Chronic obstructive pulmonary disease (COPD) is a leading cause of hospitalization and mortality globally, placing a substantial burden on healthcare systems. In Saudi Arabia, COPD admissions are rising due to demographic shifts and environmental exposures. Accurate prediction of COPD-related hospitalizations is essential for timely intervention and resource planning. This study applied machine learning (ML) techniques to predict COPD admissions using routine hospital data from major healthcare facilities in Riyadh.

Methods

A cross-sectional analysis was conducted using 41,544 patient admission records from eight major hospitals in Saudi Arabia between 2022 and 2024. The dataset included demographic, clinical, and healthcare utilization variables. Several ML classifiers: Logistic Regression, Support Vector Machine, K-Nearest Neighbors, Decision Tree, Random Forest, Gradient Boosting, and XGBoost, were developed and evaluated. The primary outcome was inpatient admission for COPD. Model performance was assessed using accuracy, precision, recall, F1-score, AUROC, and confusion matrices. SHapley Additive exPlanations (SHAP) were used to interpret model outputs and rank feature importance.

Results

The Random Forest model outperformed other classifiers with an accuracy of 0.73, precision of 0.70, recall of 0.79, F1-score of 0.74, and AUROC of 0.79. Key predictors identified by SHAP analysis included hospital name, admission count, comorbid conditions, and disease severity. Features such as gender and seasonal variation showed minimal influence on prediction outcomes. SHAP visualizations provided interpretable insights into individual-level risk contributions.

Conclusion

Machine learning models, particularly Random Forest, demonstrated moderate but promising capacity for predicting COPD admissions using routine hospital data. Model interpretability through SHAP enhances clinical relevance and supports early identification of high-risk patients. Integration of these tools into hospital systems may facilitate proactive care and improve resource allocation for respiratory conditions.

慢性阻塞性肺疾病（COPD）是全球住院和死亡的主要原因，给卫生保健系统带来了沉重负担。在沙特阿拉伯，由于人口变化和环境暴露，慢性阻塞性肺病入院人数正在上升。准确预测copd相关住院对及时干预和资源规划至关重要。本研究利用利雅得主要医疗机构的常规医院数据，应用机器学习（ML）技术预测COPD入院情况。方法对沙特阿拉伯8家主要医院2022 - 2024年间41544例住院患者进行横断面分析。数据集包括人口统计、临床和医疗保健利用变量。开发并评估了几个ML分类器：逻辑回归、支持向量机、k近邻、决策树、随机森林、梯度增强和XGBoost。主要终点为慢性阻塞性肺病住院。使用准确性、精密度、召回率、f1评分、AUROC和混淆矩阵评估模型性能。SHapley加性解释（SHAP）用于解释模型输出并对特征重要性进行排序。结果随机森林模型的准确率为0.73，精密度为0.70，召回率为0.79，f1得分为0.74，AUROC为0.79，优于其他分类器。SHAP分析确定的关键预测因素包括医院名称、入院人数、合并症和疾病严重程度。性别和季节变化等特征对预测结果的影响最小。SHAP可视化为个人层面的风险贡献提供了可解释的见解。机器学习模型，特别是随机森林模型，在使用常规医院数据预测COPD入院情况方面表现出中等但有希望的能力。通过SHAP模型的可解释性提高了临床相关性，并支持早期识别高危患者。将这些工具整合到医院系统中可以促进主动护理并改善呼吸系统疾病的资源分配。

{"title":"Predicting COPD admissions using machine learning and SHAP: An exploratory multi-hospital study in Riyadh, Saudi Arabia","authors":"Anas Ali Alhur , Jamilu Sani , Mohamed Mustaf Ahmed","doi":"10.1016/j.ibmed.2025.100312","DOIUrl":"10.1016/j.ibmed.2025.100312","url":null,"abstract":"<div><h3>Background</h3><div>Chronic obstructive pulmonary disease (COPD) is a leading cause of hospitalization and mortality globally, placing a substantial burden on healthcare systems. In Saudi Arabia, COPD admissions are rising due to demographic shifts and environmental exposures. Accurate prediction of COPD-related hospitalizations is essential for timely intervention and resource planning. This study applied machine learning (ML) techniques to predict COPD admissions using routine hospital data from major healthcare facilities in Riyadh.</div></div><div><h3>Methods</h3><div>A cross-sectional analysis was conducted using 41,544 patient admission records from eight major hospitals in Saudi Arabia between 2022 and 2024. The dataset included demographic, clinical, and healthcare utilization variables. Several ML classifiers: Logistic Regression, Support Vector Machine, K-Nearest Neighbors, Decision Tree, Random Forest, Gradient Boosting, and XGBoost, were developed and evaluated. The primary outcome was inpatient admission for COPD. Model performance was assessed using accuracy, precision, recall, F1-score, AUROC, and confusion matrices. SHapley Additive exPlanations (SHAP) were used to interpret model outputs and rank feature importance.</div></div><div><h3>Results</h3><div>The Random Forest model outperformed other classifiers with an accuracy of 0.73, precision of 0.70, recall of 0.79, F1-score of 0.74, and AUROC of 0.79. Key predictors identified by SHAP analysis included hospital name, admission count, comorbid conditions, and disease severity. Features such as gender and seasonal variation showed minimal influence on prediction outcomes. SHAP visualizations provided interpretable insights into individual-level risk contributions.</div></div><div><h3>Conclusion</h3><div>Machine learning models, particularly Random Forest, demonstrated moderate but promising capacity for predicting COPD admissions using routine hospital data. Model interpretability through SHAP enhances clinical relevance and supports early identification of high-risk patients. Integration of these tools into hospital systems may facilitate proactive care and improve resource allocation for respiratory conditions.</div></div>","PeriodicalId":73399,"journal":{"name":"Intelligence-based medicine","volume":"12 ","pages":"Article 100312"},"PeriodicalIF":0.0,"publicationDate":"2025-01-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145519505","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Optimizing ResNet50 performance using stochastic gradient descent on MRI images for Alzheimer's disease classification 在MRI图像上使用随机梯度下降优化ResNet50性能，用于阿尔茨海默病分类

Intelligence-based medicine

Pub Date : 2025-01-01 DOI: 10.1016/j.ibmed.2025.100219

Mohamed Amine Mahjoubi , Driss Lamrani , Shawki Saleh , Wassima Moutaouakil , Asmae Ouhmida , Soufiane Hamida , Bouchaib Cherradi , Abdelhadi Raihani

The field of optimization is focused on the formulation, analysis, and resolution of problems involving the minimization or maximization of functions. A particular subclass of optimization problems, known as empirical risk minimization, involves fitting a model to observed data. These problems play a central role in various areas such as machine learning, statistical modeling, and decision theory, where the objective is to find a model that best approximates underlying patterns in the data by minimizing a specified loss or risk function. In deep learning (DL) systems, various optimization algorithms are utilized with the gradient descent (GD) algorithm being one of the most significant and effective. Research studies have improved the GD algorithm and developed various successful variants, including stochastic gradient descent (SGD) with momentum, AdaGrad, RMSProp, and Adam. This article provides a comparative analysis of these stochastic gradient descent algorithms based on their accuracy, loss, and training time, as well as the loss of each algorithm in generating an optimization solution. Experiments were conducted using Transfer Learning (TL) technique based on the pre-trained ResNet50 base model for image classification, with a focus on stochastic gradient (SG) for performances optimization. The case study in this work is based on a data extract from the Alzheimer's image dataset, which contains four classes such as Mild Demented, Moderate Demented, Non-Demented, and Very Mild Demented. The obtained results with the Adam and SGD momentum optimizers achieved the highest accuracy of 97.66 % and 97.58 %, respectively.

优化领域的重点是制定、分析和解决涉及函数的最小化或最大化的问题。优化问题的一个特殊子类，被称为经验风险最小化，涉及到将模型拟合到观测数据。这些问题在机器学习、统计建模和决策理论等各个领域发挥着核心作用，这些领域的目标是通过最小化指定的损失或风险函数来找到最接近数据中潜在模式的模型。在深度学习（DL）系统中，有各种各样的优化算法，其中梯度下降（GD）算法是最重要和最有效的算法之一。研究改进了GD算法，并开发了各种成功的变体，包括带动量的随机梯度下降（SGD）、AdaGrad、RMSProp和Adam。本文根据这些随机梯度下降算法的精度、损失和训练时间，以及每种算法在生成优化解时的损失，对它们进行了比较分析。基于预训练的ResNet50基模型，采用迁移学习（TL）技术进行图像分类实验，重点采用随机梯度（SG）进行性能优化。本研究的案例研究基于阿尔茨海默病图像数据集的数据提取，该数据集包含轻度痴呆、中度痴呆、非痴呆和极轻度痴呆等四类。使用Adam动量优化器和SGD动量优化器获得的结果准确率最高，分别为97.66%和97.58%。

{"title":"Optimizing ResNet50 performance using stochastic gradient descent on MRI images for Alzheimer's disease classification","authors":"Mohamed Amine Mahjoubi , Driss Lamrani , Shawki Saleh , Wassima Moutaouakil , Asmae Ouhmida , Soufiane Hamida , Bouchaib Cherradi , Abdelhadi Raihani","doi":"10.1016/j.ibmed.2025.100219","DOIUrl":"10.1016/j.ibmed.2025.100219","url":null,"abstract":"<div><div>The field of optimization is focused on the formulation, analysis, and resolution of problems involving the minimization or maximization of functions. A particular subclass of optimization problems, known as empirical risk minimization, involves fitting a model to observed data. These problems play a central role in various areas such as machine learning, statistical modeling, and decision theory, where the objective is to find a model that best approximates underlying patterns in the data by minimizing a specified loss or risk function. In deep learning (DL) systems, various optimization algorithms are utilized with the gradient descent (GD) algorithm being one of the most significant and effective. Research studies have improved the GD algorithm and developed various successful variants, including stochastic gradient descent (SGD) with momentum, AdaGrad, RMSProp, and Adam. This article provides a comparative analysis of these stochastic gradient descent algorithms based on their accuracy, loss, and training time, as well as the loss of each algorithm in generating an optimization solution. Experiments were conducted using Transfer Learning (TL) technique based on the pre-trained ResNet50 base model for image classification, with a focus on stochastic gradient (SG) for performances optimization. The case study in this work is based on a data extract from the Alzheimer's image dataset, which contains four classes such as Mild Demented, Moderate Demented, Non-Demented, and Very Mild Demented. The obtained results with the Adam and SGD momentum optimizers achieved the highest accuracy of 97.66 % and 97.58 %, respectively.</div></div>","PeriodicalId":73399,"journal":{"name":"Intelligence-based medicine","volume":"11 ","pages":"Article 100219"},"PeriodicalIF":0.0,"publicationDate":"2025-01-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143173637","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

A mobile application LukaKu as a tool for detecting external wounds with artificial intelligence 一个移动应用程序LukaKu作为人工智能检测外部伤口的工具

Intelligence-based medicine

Pub Date : 2025-01-01 DOI: 10.1016/j.ibmed.2025.100200

Dessy Novita , Herika Hayurani , Eva Krishna Sutedja , Firdaus Ryan Pratomo , Achmad Dino Saputra , Zahra Ramadhanti , Nuryadin Abutani , Muhammad Rafi Triandi , Aldin Mubarok Guferol , Anindya Apriliyanti Pravitasari , Fajar Wira Adikusuma , Atiek Rostika Noviyanti

This study was conducted due to the lack of applications that can assist people intreating common external wounds. Therefore, we proposed the application of image-based detection which takes external wounds and identifies them using Artificial Intelligence namely LukaKu. In addition to detecting the type of wound that occurs, the application is expected to be able to produce first aid and medicine for each existing external wound label. The model used is YOLOv5 with various versions, namely YOLOv5n, YOLOv5s, YOLOv5m, YOLOv5l, and YOLOv5x. By calculating the validation data, each version has its own precision, recall, f1-score, and Mean Average Precision (mAP) values which are the comparison factors in determining the best model version, where YOLOv5l with mAP value of 0.785 is the best result and YOLOv5n with mAP value of 0.588 is the result with the lowest value. In the model development process, datasets of external injuries are needed to be used during the training process and test datasets for each existing model version. After each version of the model has been successfully built and analysed, the model with the best value is implemented in the mobile application, making it easier for users to access.

这项研究是由于缺乏应用程序，可以帮助人们治疗常见的外部伤口。因此，我们提出了基于图像的检测应用，该检测采用人工智能即LukaKu来识别外部伤口。除了检测发生的伤口类型之外，该应用程序预计能够为每个现有的外部伤口标签生产急救和药物。型号为YOLOv5，有YOLOv5n、YOLOv5s、YOLOv5m、YOLOv5l、YOLOv5x等多个版本。通过计算验证数据，每个版本都有自己的精度、召回率、f1-score和Mean Average precision （mAP）值，这些值是确定最佳模型版本的比较因素，其中mAP值为0.785的YOLOv5l为最佳结果，mAP值为0.588的YOLOv5n为最低结果。在模型开发过程中，在训练过程中需要使用外伤性数据集，在现有的各个模型版本中需要使用测试数据集。在成功构建和分析了每个版本的模型后，将最有价值的模型实现在移动应用程序中，使用户更容易访问。

{"title":"A mobile application LukaKu as a tool for detecting external wounds with artificial intelligence","authors":"Dessy Novita , Herika Hayurani , Eva Krishna Sutedja , Firdaus Ryan Pratomo , Achmad Dino Saputra , Zahra Ramadhanti , Nuryadin Abutani , Muhammad Rafi Triandi , Aldin Mubarok Guferol , Anindya Apriliyanti Pravitasari , Fajar Wira Adikusuma , Atiek Rostika Noviyanti","doi":"10.1016/j.ibmed.2025.100200","DOIUrl":"10.1016/j.ibmed.2025.100200","url":null,"abstract":"<div><div>This study was conducted due to the lack of applications that can assist people intreating common external wounds. Therefore, we proposed the application of image-based detection which takes external wounds and identifies them using Artificial Intelligence namely LukaKu. In addition to detecting the type of wound that occurs, the application is expected to be able to produce first aid and medicine for each existing external wound label. The model used is YOLOv5 with various versions, namely YOLOv5n, YOLOv5s, YOLOv5m, YOLOv5l, and YOLOv5x. By calculating the validation data, each version has its own precision, recall, f1-score, and Mean Average Precision (mAP) values which are the comparison factors in determining the best model version, where YOLOv5l with mAP value of 0.785 is the best result and YOLOv5n with mAP value of 0.588 is the result with the lowest value. In the model development process, datasets of external injuries are needed to be used during the training process and test datasets for each existing model version. After each version of the model has been successfully built and analysed, the model with the best value is implemented in the mobile application, making it easier for users to access.</div></div>","PeriodicalId":73399,"journal":{"name":"Intelligence-based medicine","volume":"11 ","pages":"Article 100200"},"PeriodicalIF":0.0,"publicationDate":"2025-01-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143174331","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Image-based machine learning model as a tool for classification of [18F]PR04.MZ PET images in patients with parkinsonian syndrome 将基于图像的机器学习模型作为帕金森综合征患者[18F]PR04.MZ PET 图像分类的工具

Intelligence-based medicine

Pub Date : 2025-01-01 DOI: 10.1016/j.ibmed.2025.100232

Maria Jiménez , Cristian Soza-Ried , Vasko Kramer , Sebastian A. Ríos , Arlette Haeger , Carlos Juri , Horacio Amaral , Pedro Chana-Cuevas

Parkinsonian syndrome (PS) is characterized by bradykinesia, resting tremor, rigidity, and encapsulates the clinical manifestation observed in various neurodegenerative disorders. Positron emission tomography (PET) imaging plays an important role in diagnosing PS by detecting the progressive loss of dopaminergic neurons. This study aimed to develop and compare five machine-learning models for the automatic classification of 204 [¹⁸F]PR04.MZ PET images, distinguishing between patients with PS and subjects without clinical evidence for dopaminergic deficit (SWEDD). Previously analyzed and classified by three expert blind readers into PS compatible (1) and SWEDDs (0), the dataset was processed in both two-dimensional and three-dimensional formats. Five widely used pattern recognition algorithms were trained and validated their performance. These algorithms were compared against the majority reading of expert diagnosis, considered the gold standard. Comparing the accuracy of 2D and 3D format images suggests that, without the depth dimension, a single image may overemphasize specific regions. Overall, three models outperformed with an accuracy greater than 98 %, demonstrating that machine-learning models trained with [¹⁸F]PR04.MZ PET images can provide a highly accurate and precise tool to support clinicians in automatic PET image analysis. This approach may be a first step in reducing the time required for interpretation, as well as increase certainty in the diagnostic process.

帕金森综合征（Parkinsonian Syndrome，PS）以运动迟缓、静止性震颤和僵直为特征，是各种神经退行性疾病的临床表现。正电子发射断层扫描（PET）成像通过检测多巴胺能神经元的逐渐丧失，在诊断帕金森综合征中发挥着重要作用。本研究旨在开发和比较五种机器学习模型，用于对204张[18F]PR04.MZ PET图像进行自动分类，区分PS患者和无多巴胺能缺失临床证据的受试者（SWEDD）。该数据集之前由三位盲人专家进行了分析和分类，分为 PS 相容性（1）和 SWEDD（0），并以二维和三维格式进行了处理。对五种广泛使用的模式识别算法进行了训练，并对其性能进行了验证。这些算法与被视为金标准的专家诊断的多数读数进行了比较。比较二维和三维格式图像的准确性表明，如果没有深度维度，单一图像可能会过分强调特定区域。总体而言，三个模型的准确率都超过了 98%，这表明使用[18F]PR04.MZ PET 图像训练的机器学习模型可以提供一种高度准确和精确的工具，为临床医生自动 PET 图像分析提供支持。这种方法可能是减少判读所需时间的第一步，并能提高诊断过程的确定性。

{"title":"Image-based machine learning model as a tool for classification of [18F]PR04.MZ PET images in patients with parkinsonian syndrome","authors":"Maria Jiménez , Cristian Soza-Ried , Vasko Kramer , Sebastian A. Ríos , Arlette Haeger , Carlos Juri , Horacio Amaral , Pedro Chana-Cuevas","doi":"10.1016/j.ibmed.2025.100232","DOIUrl":"10.1016/j.ibmed.2025.100232","url":null,"abstract":"<div><div>Parkinsonian syndrome (PS) is characterized by bradykinesia, resting tremor, rigidity, and encapsulates the clinical manifestation observed in various neurodegenerative disorders. Positron emission tomography (PET) imaging plays an important role in diagnosing PS by detecting the progressive loss of dopaminergic neurons. This study aimed to develop and compare five machine-learning models for the automatic classification of 204 [<sup>18</sup>F]PR04.MZ PET images, distinguishing between patients with PS and subjects without clinical evidence for dopaminergic deficit (SWEDD). Previously analyzed and classified by three expert blind readers into PS compatible (1) and SWEDDs (0), the dataset was processed in both two-dimensional and three-dimensional formats. Five widely used pattern recognition algorithms were trained and validated their performance. These algorithms were compared against the majority reading of expert diagnosis, considered the gold standard. Comparing the accuracy of 2D and 3D format images suggests that, without the depth dimension, a single image may overemphasize specific regions. Overall, three models outperformed with an accuracy greater than 98 %, demonstrating that machine-learning models trained with [<sup>18</sup>F]PR04.MZ PET images can provide a highly accurate and precise tool to support clinicians in automatic PET image analysis. This approach may be a first step in reducing the time required for interpretation, as well as increase certainty in the diagnostic process.</div></div>","PeriodicalId":73399,"journal":{"name":"Intelligence-based medicine","volume":"11 ","pages":"Article 100232"},"PeriodicalIF":0.0,"publicationDate":"2025-01-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143628516","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

A comparison of techniques for predicting telehealth visit failure 预测远程医疗访问失败的技术比较

Intelligence-based medicine

Pub Date : 2025-01-01 DOI: 10.1016/j.ibmed.2025.100235

Alexander J. Idarraga , David F. Schneider

Objective

Telehealth is an increasingly important method for delivering care. Health systems lack the ability to accurately predict which telehealth visits will fail due to poor connection, poor technical literacy, or other reasons. This results in wasted resources and disrupted patient care. The purpose of this study is to characterize and compare various methods for predicting telehealth visit failure, and to determine the prediction method most suited for implementation in a real-time operational setting.

Methods

A single-center, retrospective cohort study was conducted using data sourced from our data warehouse. Patient demographic information and data characterizing prior visit success and engagement with electronic health tools were included. Three main model types were evaluated: an existing scoring model developed by Hughes et al., a regression-based scoring model, and Machine Learning classifiers. Variables were selected for their importance and anticipated availability; Number Needed to Treat was used to demonstrate the number of interventions (e.g. pre-visit phone calls) required to improve success rates in the context of weekly patient volumes.

Results

217, 229 visits spanning 480 days were evaluated, of which 22,443 (10.33 %) met criteria for failure. Hughes et al.’s model applied to our data yielded an Area Under the Receiver Operating Characteristics Curve (AUC ROC) of 0.678 when predicting failure. A score-based model achieved an AUC ROC of 0.698. Logistic Regression, Random Forest, and Gradient Boosting models demonstrated AUC ROCs ranging from 0.7877 to 0.7969. A NNT of 32 was achieved if the 263 highest-risk patients were selected in a low-volume week using the RF classifier, compared to an expected NNT of 90 if the same number of patients were randomly selected.

Conclusions

Machine Learning classifiers demonstrated superiority over score-based methods for predicting telehealth visit failure. Prospective evaluation is required; evaluation using NNT as a metric can help to operationalize these models.

目的远程医疗是一种日益重要的医疗服务方式。卫生系统缺乏准确预测哪些远程医疗访问将由于连接不良、技术素养低下或其他原因而失败的能力。这导致资源浪费和病人护理中断。本研究的目的是描述和比较各种预测远程医疗访问失败的方法，并确定最适合在实时操作环境中实施的预测方法。方法采用单中心、回顾性队列研究，数据来源于我们的数据仓库。包括患者人口统计信息和表征先前访问成功和使用电子健康工具的数据。评估了三种主要的模型类型：Hughes等人开发的现有评分模型，基于回归的评分模型和机器学习分类器。根据变量的重要性和预期可用性选择变量；需要治疗的人数用于展示在每周患者数量的情况下提高成功率所需的干预措施的数量（例如，就诊前电话）。结果共评估就诊217229次，共计480 d，其中22443次（10.33%）符合不合格标准。Hughes等人的模型应用于我们的数据，在预测失败时，接受者工作特征曲线下面积（AUC ROC）为0.678。基于评分的模型的AUC ROC为0.698。Logistic回归、随机森林和梯度增强模型的AUC roc范围为0.7877 ~ 0.7969。如果在低容量周内使用RF分类器选择263名风险最高的患者，则NNT为32，而如果随机选择相同数量的患者，则NNT为90。结论机器学习分类器在预测远程医疗就诊失败方面优于基于分数的方法。需要前瞻性评价；使用NNT作为度量的评估可以帮助这些模型的操作。

{"title":"A comparison of techniques for predicting telehealth visit failure","authors":"Alexander J. Idarraga , David F. Schneider","doi":"10.1016/j.ibmed.2025.100235","DOIUrl":"10.1016/j.ibmed.2025.100235","url":null,"abstract":"<div><h3>Objective</h3><div>Telehealth is an increasingly important method for delivering care. Health systems lack the ability to accurately predict which telehealth visits will fail due to poor connection, poor technical literacy, or other reasons. This results in wasted resources and disrupted patient care. The purpose of this study is to characterize and compare various methods for predicting telehealth visit failure, and to determine the prediction method most suited for implementation in a real-time operational setting.</div></div><div><h3>Methods</h3><div>A single-center, retrospective cohort study was conducted using data sourced from our data warehouse. Patient demographic information and data characterizing prior visit success and engagement with electronic health tools were included. Three main model types were evaluated: an existing scoring model developed by Hughes et al., a regression-based scoring model, and Machine Learning classifiers. Variables were selected for their importance and anticipated availability; Number Needed to Treat was used to demonstrate the number of interventions (e.g. pre-visit phone calls) required to improve success rates in the context of weekly patient volumes.</div></div><div><h3>Results</h3><div>217, 229 visits spanning 480 days were evaluated, of which 22,443 (10.33 %) met criteria for failure. Hughes et al.’s model applied to our data yielded an Area Under the Receiver Operating Characteristics Curve (AUC ROC) of 0.678 when predicting failure. A score-based model achieved an AUC ROC of 0.698. Logistic Regression, Random Forest, and Gradient Boosting models demonstrated AUC ROCs ranging from 0.7877 to 0.7969. A NNT of 32 was achieved if the 263 highest-risk patients were selected in a low-volume week using the RF classifier, compared to an expected NNT of 90 if the same number of patients were randomly selected.</div></div><div><h3>Conclusions</h3><div>Machine Learning classifiers demonstrated superiority over score-based methods for predicting telehealth visit failure. Prospective evaluation is required; evaluation using NNT as a metric can help to operationalize these models.</div></div>","PeriodicalId":73399,"journal":{"name":"Intelligence-based medicine","volume":"11 ","pages":"Article 100235"},"PeriodicalIF":0.0,"publicationDate":"2025-01-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143747628","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Comparative analysis of deep learning and machine learning techniques for forecasting new malaria cases in Cameroon’s Adamaoua region 深度学习和机器学习技术在喀麦隆阿达马乌阿地区预测新疟疾病例的比较分析

Intelligence-based medicine

Pub Date : 2025-01-01 DOI: 10.1016/j.ibmed.2025.100220

Esaie Naroum , Ebenezer Maka Maka , Hamadjam Abboubakar , Paul Dayang , Appolinaire Batoure Bamana , Benjamin Garga , Hassana Daouda Daouda , Mohsen Bakouri , Ilyas Khan

The Plasmodium parasite, which causes malaria is transmitted by Anopheles mosquitoes, and remains a major development barrier in Africa. This is particularly true considering the conducive environment that promotes the spread of malaria. This study examines several machine learning approaches, such as long short term memory (LSTM), random forests (RF), support vector machines (SVM), and data regularization models including Ridge, Lasso, and ElasticNet, in order to forecast the occurrence of malaria in the Adamaoua region of Cameroon. The LSTM, a recurrent neural network variant, performed the best with 76% accuracy and a low error rate (RMSE = 0.08). Statistical evidence indicates that temperatures exceeding 34 degrees halt mosquito vector reproduction, thereby slowing the spread of malaria. However, humidity increases the morbidity of the condition. The survey also identified high-risk areas in Ngaoundéré Rural and Urban and Meiganga. Between 2018 and 2022, the Adamaoua region had 20.1%, 12.3%, and 10.0% of malaria cases, respectively, in these locations. According to the estimate, the number of malaria cases in the Adamaoua region will rise gradually between 2023 and 2026, peaking in 2029 before declining in 2031.

引起疟疾的疟原虫是由按蚊传播的，它仍然是非洲的一个主要发展障碍。考虑到促进疟疾传播的有利环境，这一点尤其正确。本研究探讨了几种机器学习方法，如长短期记忆（LSTM）、随机森林（RF）、支持向量机（SVM）和数据正则化模型（包括Ridge、Lasso和ElasticNet），以预测喀麦隆阿达马瓦地区疟疾的发生。LSTM，一种循环神经网络变体，表现最好，准确率为76%，错误率低（RMSE = 0.08）。统计证据表明，超过34度的温度会阻止蚊子媒介的繁殖，从而减缓疟疾的传播。然而，湿度增加了病情的发病率。调查还确定了恩oundd农村和城市以及梅甘加的高风险地区。2018年至2022年期间，阿达马乌瓦地区分别占这些地区疟疾病例的20.1%、12.3%和10.0%。据估计，阿达马乌瓦地区的疟疾病例数将在2023年至2026年期间逐步上升，在2029年达到峰值，然后在2031年下降。

{"title":"Comparative analysis of deep learning and machine learning techniques for forecasting new malaria cases in Cameroon’s Adamaoua region","authors":"Esaie Naroum , Ebenezer Maka Maka , Hamadjam Abboubakar , Paul Dayang , Appolinaire Batoure Bamana , Benjamin Garga , Hassana Daouda Daouda , Mohsen Bakouri , Ilyas Khan","doi":"10.1016/j.ibmed.2025.100220","DOIUrl":"10.1016/j.ibmed.2025.100220","url":null,"abstract":"<div><div>The Plasmodium parasite, which causes malaria is transmitted by Anopheles mosquitoes, and remains a major development barrier in Africa. This is particularly true considering the conducive environment that promotes the spread of malaria. This study examines several machine learning approaches, such as long short term memory (LSTM), random forests (RF), support vector machines (SVM), and data regularization models including Ridge, Lasso, and ElasticNet, in order to forecast the occurrence of malaria in the Adamaoua region of Cameroon. The LSTM, a recurrent neural network variant, performed the best with 76% accuracy and a low error rate (RMSE = 0.08). Statistical evidence indicates that temperatures exceeding 34 degrees halt mosquito vector reproduction, thereby slowing the spread of malaria. However, humidity increases the morbidity of the condition. The survey also identified high-risk areas in Ngaoundéré Rural and Urban and Meiganga. Between 2018 and 2022, the Adamaoua region had 20.1%, 12.3%, and 10.0% of malaria cases, respectively, in these locations. According to the estimate, the number of malaria cases in the Adamaoua region will rise gradually between 2023 and 2026, peaking in 2029 before declining in 2031.</div></div>","PeriodicalId":73399,"journal":{"name":"Intelligence-based medicine","volume":"11 ","pages":"Article 100220"},"PeriodicalIF":0.0,"publicationDate":"2025-01-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143388305","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Optimizing breast cancer diagnosis with convolutional autoencoders: Enhanced performance through modified loss functions 用卷积自编码器优化乳腺癌诊断：通过修改损失函数增强性能

Intelligence-based medicine

Pub Date : 2025-01-01 DOI: 10.1016/j.ibmed.2025.100248

ArunaDevi Karuppasamy , Hamza zidoum , Majda Said Sultan Al-Rashdi , Maiya Al-Bahri

The Deep Learning (DL) has demonstrated a significant impact on a various pattern recognition applications, resulting in significant advancements in areas such as visual recognition, autonomous cars, language processing, and healthcare. Nowadays, deep learning was widely applied on the medical images to identify the diseases efficiently. Still, the use of applications in clinical settings is now limited to a small number. The main factors to this might be due to an inadequate annotated data, noises in the images and challenges related to collecting data. Our research proposed a convolutional autoencoder to classify the breast cancer tumors, using the Sultan Qaboos University Hospital(SQUH) and BreakHis datasets. The proposed model named Convolutional AutoEncoder with modified Loss Function (CAE-LF) achieved a good performance, by attaining a F1-score of 0.90, recall of 0.89, and accuracy of 91%. The results obtained are comparable to those obtained in earlier researches. Additional analyses conducted on the SQUH dataset demonstrate that it yields a good performance with an F1-score of 0.91, 0.93, 0.92, and 0.93 for 4x, 10x, 20x, and 40x magnifications, respectively. Our study highlights the potential of deep learning in analyzing medical images to classify breast tumors.

深度学习（DL）已经对各种模式识别应用产生了重大影响，在视觉识别、自动驾驶汽车、语言处理和医疗保健等领域取得了重大进展。目前，深度学习被广泛应用于医学图像，以有效地识别疾病。尽管如此，应用程序在临床环境中的使用现在仅限于少数。造成这种情况的主要因素可能是由于注释数据不足，图像中的噪声以及与收集数据相关的挑战。我们的研究提出了一种卷积自编码器来分类乳腺癌肿瘤，使用苏丹卡布斯大学医院（SQUH）和BreakHis数据集。所提出的基于改进损失函数的卷积自编码器（CAE-LF）模型取得了良好的性能，f1得分为0.90，召回率为0.89，准确率为91%。所得结果与早期的研究结果相当。对SQUH数据集进行的进一步分析表明，在4倍、10倍、20倍和40倍的放大倍数下，它的f1得分分别为0.91、0.93、0.92和0.93，表现良好。我们的研究强调了深度学习在分析医学图像以分类乳腺肿瘤方面的潜力。

{"title":"Optimizing breast cancer diagnosis with convolutional autoencoders: Enhanced performance through modified loss functions","authors":"ArunaDevi Karuppasamy , Hamza zidoum , Majda Said Sultan Al-Rashdi , Maiya Al-Bahri","doi":"10.1016/j.ibmed.2025.100248","DOIUrl":"10.1016/j.ibmed.2025.100248","url":null,"abstract":"<div><div>The Deep Learning (DL) has demonstrated a significant impact on a various pattern recognition applications, resulting in significant advancements in areas such as visual recognition, autonomous cars, language processing, and healthcare. Nowadays, deep learning was widely applied on the medical images to identify the diseases efficiently. Still, the use of applications in clinical settings is now limited to a small number. The main factors to this might be due to an inadequate annotated data, noises in the images and challenges related to collecting data. Our research proposed a convolutional autoencoder to classify the breast cancer tumors, using the Sultan Qaboos University Hospital(SQUH) and BreakHis datasets. The proposed model named Convolutional AutoEncoder with modified Loss Function (CAE-LF) achieved a good performance, by attaining a F1-score of 0.90, recall of 0.89, and accuracy of 91%. The results obtained are comparable to those obtained in earlier researches. Additional analyses conducted on the SQUH dataset demonstrate that it yields a good performance with an F1-score of 0.91, 0.93, 0.92, and 0.93 for 4x, 10x, 20x, and 40x magnifications, respectively. Our study highlights the potential of deep learning in analyzing medical images to classify breast tumors.</div></div>","PeriodicalId":73399,"journal":{"name":"Intelligence-based medicine","volume":"11 ","pages":"Article 100248"},"PeriodicalIF":0.0,"publicationDate":"2025-01-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143887937","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0