Emegano Declan Ikechukwu, Mustapha Mubarak Taiwo, Ozsahin Ilker, Ozsahin Dilber Uzun, Uzun Berna
Operational Research Centre in Healthcare, Near East University, TRNC Mersin 10, Nicosia 99138, Turkey.
Department of Medical Diagnostic Imaging, College of Health Sciences, University of Sharjah, Sharjah 27272, United Arab Emirates.
Bioengineering (Basel). 2025 Apr 1;12(4):369. doi: 10.3390/bioengineering12040369.
Prostate cancer is a leading cause of cancer-related morbidity and mortality worldwide, with diagnostic challenges magnified in underrepresented regions like sub-Saharan Africa. This study introduces a novel application of ConvNeXt, an advanced convolutional neural network architecture, for multi-class classification of prostate histopathological images into normal, benign, and malignant categories. The dataset, sourced from a tertiary healthcare institution in Nigeria, represents a typically underserved African population, addressing critical disparities in global diagnostic research. We also used the ProstateX dataset (2017) from The Cancer Imaging Archive (TCIA) to validate our result. A comprehensive pipeline was developed, leveraging advanced data augmentation, Grad-CAM for interpretability, and an ablation study to enhance model optimization and robustness. The ConvNeXt model achieved an accuracy of 98%, surpassing the performance of traditional CNNs (ResNet50, 93%; EfficientNet, 94%; DenseNet, 92%) and transformer-based models (ViT, 88%; CaiT, 86%; Swin Transformer, 95%; RegNet, 94%). Also, using the ProstateX dataset, the ConvNeXt model recorded 87.2%, 85.7%, 86.4%, and 0.92 as accuracy, recall, F1 score, and AUC, respectively, as validation results. Its hybrid architecture combines the strengths of CNNs and transformers, enabling superior feature extraction. Grad-CAM visualizations further enhance explainability, bridging the gap between computational predictions and clinical trust. Ablation studies demonstrated the contributions of data augmentation, optimizer selection, and learning rate tuning to model performance, highlighting its robustness and adaptability for deployment in low-resource settings. This study advances equitable health care by addressing the lack of regional representation in diagnostic datasets and employing a clinically aligned three-class classification approach. Combining high performance, interpretability, and scalability, this work establishes a foundation for future research on diverse and underrepresented populations, fostering global inclusivity in cancer diagnostics.
前列腺癌是全球癌症相关发病和死亡的主要原因，在撒哈拉以南非洲等代表性不足的地区，诊断挑战更加严峻。本研究介绍了一种先进的卷积神经网络架构ConvNeXt在前列腺组织病理学图像多分类中的新应用，该分类可分为正常、良性和恶性类别。该数据集来自尼日利亚的一家三级医疗机构，代表了典型的医疗服务不足的非洲人群，解决了全球诊断研究中的关键差异问题。我们还使用了癌症成像存档（TCIA）的ProstateX数据集（2017年）来验证我们的结果。我们开发了一个综合流程，利用先进的数据增强、用于可解释性的Grad-CAM以及消融研究来提高模型的优化和鲁棒性。ConvNeXt模型的准确率达到了98%，超过了传统卷积神经网络（ResNet50为93%；EfficientNet为94%；DenseNet为92%）和基于Transformer的模型（ViT为88%；CaiT为86%；Swin Transformer为95%；RegNet为94%）的性能。此外，使用ProstateX数据集，ConvNeXt模型的验证结果分别为准确率87.2%、召回率85.7%、F1分数86.4%和AUC为0.92。其混合架构结合了卷积神经网络和Transformer的优势，实现了卓越的特征提取。Grad-CAM可视化进一步增强了可解释性，弥合了计算预测与临床信任之间的差距。消融研究证明了数据增强、优化器选择和学习率调整对模型性能的贡献，突出了其在低资源环境中部署的鲁棒性和适应性。本研究通过解决诊断数据集中缺乏区域代表性的问题，并采用临床一致的三类分类方法，推动了公平医疗。这项工作结合了高性能、可解释性和可扩展性，为未来针对多样化和代表性不足人群的研究奠定了基础，促进了癌症诊断的全球包容性。
