Hanif Muhammad Shehzad, Bilal Muhammad, Alsaggaf Abdullah H, Al-Saggaf Ubaid M
Department of Electrical and Computer Engineering, King Abdulaziz University, Jeddah 21589, Saudi Arabia.
Unit of Allergy and Immunology, Department of Pediatrics, King Abdulaziz University, Jeddah 21589, Saudi Arabia.
Bioengineering (Basel). 2025 May 31;12(6):593. doi: 10.3390/bioengineering12060593.
This article addresses the non-trivial problem of classifying thoracic diseases in chest X-ray (CXR) images. A single CXR image may exhibit multiple diseases, making this a multi-label classification problem. Additionally, the inherent class imbalance makes the task even more challenging as some diseases occur more frequently than others. Our methodology is based on transfer learning aiming to fine-tune a pretrained DenseNet121 model using CXR images from the NIH Chest X-ray14 dataset. Training from scratch would require a large-scale dataset containing millions of images, which is not available in the public domain for this multi-label classification task. To address class imbalance problem, we propose a rank-based loss derived from the Zero-bounded Log-sum-exp and Pairwise Rank-based (ZLPR) loss, which we refer to as focal ZLPR (FZLPR). In designing FZLPR, we draw inspiration from the focal loss where the objective is to emphasize hard-to-classify examples (instances of rare diseases) during training compared to well-classified ones. We achieve this by incorporating a "temperature" parameter to scale the label scores predicted by the model during training in the original ZLPR loss function. Experimental results on the NIH Chest X-ray14 dataset demonstrate that FZLPR loss outperforms other loss functions including binary cross entropy (BCE) and focal loss. Moreover, by using test-time augmentations, our model trained using FZLPR loss achieves an average AUC of 80.96% which is competitive with existing approaches.
本文探讨了胸部X光（CXR）图像中胸部疾病分类这一重要问题。一张CXR图像可能呈现多种疾病，这使其成为一个多标签分类问题。此外，固有的类别不平衡使得该任务更具挑战性，因为某些疾病比其他疾病出现得更频繁。我们的方法基于迁移学习，旨在使用来自美国国立卫生研究院（NIH）胸部X光14数据集的CXR图像对预训练的DenseNet121模型进行微调。从头开始训练需要一个包含数百万张图像的大规模数据集，而在公共领域中没有用于此多标签分类任务的此类数据集。为了解决类别不平衡问题，我们提出了一种基于排序的损失函数，它源自零边界对数和指数函数以及基于成对排序的（ZLPR）损失函数，我们将其称为焦点ZLPR（FZLPR）。在设计FZLPR时，我们从焦点损失函数中获得灵感，其目标是在训练过程中相较于分类良好的示例（常见疾病的实例），强调难以分类的示例（罕见疾病的实例）。我们通过在原始ZLPR损失函数中引入一个“温度”参数来缩放模型在训练期间预测的标签分数来实现这一点。在NIH胸部X光14数据集上的实验结果表明，FZLPR损失函数优于其他损失函数，包括二元交叉熵（BCE）和焦点损失函数。此外，通过使用测试时增强技术，我们使用FZLPR损失函数训练的模型实现了80.96%的平均AUC，与现有方法具有竞争力。
