Chest CT and AI Predict Breast Cancer and Lymph Node Spread

New Study Explores Deep Learning to Improve Breast Cancer Diagnosis Using Routine CT Scans

Breast cancer remains the most common malignancy among women globally and is a major cause of cancer-related death. According to the American Cancer Society, nearly 320,000 people in the United States will be diagnosed with breast cancer by 2025, and over 42,000 will die from it. Early detection plays a crucial role in improving survival. However, distinguishing between early-stage breast cancer and benign breast masses can be challenging. Additionally, identifying whether cancer has spread to the axillary lymph nodes (ALNM) is essential for planning treatment and assessing prognosis.

Traditional imaging methods such as mammography, ultrasound (US), magnetic resonance imaging (MRI), and positron emission tomography-computed tomography (PET-CT) are commonly used, but they come with limitations. These include patient discomfort, long examination times, high costs, and difficulty obtaining a clear diagnosis with a single technique.

In contrast, non-contrast chest computed tomography (CT), which is widely used for evaluating lung, cardiac, and mediastinal conditions, is increasingly available due to routine hospital protocols and physical exams. Since chest CT typically includes the entire breast region, researchers have begun exploring its potential for breast cancer detection—especially using artificial intelligence (AI).

Study Objective: A Dual-Purpose Deep Learning Model

A team of researchers in Shandong, China, led by Jingxiang Sun and colleagues, developed a deep learning model based on non-contrast chest CT images. The goal: to predict both whether breast masses are benign or malignant, and whether early-stage breast cancer has spread to axillary lymph nodes.

Their retrospective study analyzed chest CT images from 482 patients with breast masses collected between 2018 and 2023. Among them, 224 cases were benign and 258 were malignant. The malignant group included 91 patients with confirmed ALNM and 167 without.

All patients underwent non-contrast CT scans prior to surgery. “This approach eliminates the need for contrast agents, reduces reliance on additional imaging modalities like MRI or PET-CT, and ultimately saves examination time and costs while minimizing unnecessary radiation exposure,” the authors explain.

Methodology: Applying ResNet Deep Learning Architectures

The researchers applied three types of Residual Network (ResNet) architectures—ResNet-34, ResNet-50, and ResNet-101—commonly used in computer vision tasks. All models were trained using images of the breast region, which were resized and normalized to ensure consistency. The dataset was split into training (80%), validation (10%), and test (10%) groups.

Each model was tasked with two classification objectives:

1. Differentiating benign from malignant breast masses.

2. Predicting ALNM in early-stage breast cancer patients.

Model performance was evaluated based on accuracy (ACC), sensitivity (SEN), specificity (SPE), positive predictive value (PPV), negative predictive value (NPV), and the area under the receiver operating characteristic curve (AUC).

Results: ResNet-101 Shows Best Performance

All models performed well, but ResNet-101 consistently outperformed the others in both tasks.

For mass classification:

• ResNet-101 achieved an AUC of 0.964, an ACC of 0.859, SEN of 0.929, and SPE of 0.917.

• The other models (ResNet-34 and ResNet-50) also demonstrated strong performance, with AUCs above 0.96.

For ALNM prediction:

• ResNet-101 again led with an AUC of 0.951, ACC of 0.874, SEN of 0.865, and SPE of 0.909.

• ResNet-50 and ResNet-34 followed, with AUCs of 0.936 and 0.819, respectively.

The study found that classification of breast masses had slightly better overall accuracy than ALNM prediction. The authors suggest this may be due to more distinct visual features in masses—such as shape, edge definition, and density—compared to the complex patterns involved in lymph node involvement.

Practical Implications: Towards Opportunistic Breast Cancer Screening

This study presents a practical opportunity: using non-contrast chest CT scans—already being performed for other health reasons—to screen for breast abnormalities. Integrating deep learning models into clinical workflows could allow radiologists to identify suspicious breast lesions or ALNM early, without the need for additional imaging or invasive procedures.

“Non-contrast chest CT offers several advantages,” the authors note. “It is painless, does not require multiple patient positions, and has higher spatial resolution than ultrasound.”

Compared to enhanced CT or PET-CT, non-contrast CT has lower radiation exposure and avoids contrast-related risks. It also provides a more accessible and cost-effective alternative to breast MRI, making it especially relevant in resource-limited settings.

Context in the Field: AI and Radiomics in Breast Imaging

While several studies have used contrast-enhanced CT (CECT) with radiomics or AI models to assess breast cancer and ALNM, few have focused on non-contrast CT or on performing both tasks in a single model. Earlier studies using radiomics achieved AUCs as high as 0.94 for ALNM prediction but required contrast agents.

The authors emphasize that their model uses non-contrast CT exclusively and handles both classification and metastasis prediction, broadening its clinical utility.

Limitations and Future Directions

Despite its strengths, the study has several limitations:

• It is retrospective and single-center.

• The dataset was imbalanced, with only early-stage cancers and a non-natural patient age distribution.

• Imaging data came from multiple CT scanners, which may introduce variability.

The authors acknowledge the need for prospective, multi-center studies to validate their findings. Future work will focus on expanding datasets and optimizing models specifically for ALNM prediction.

Conclusion: A Step Toward Accessible AI-Powered Diagnostics

This study demonstrates the potential of deep learning models based on non-contrast chest CT scans to identify breast cancer and assess lymph node involvement accurately. These models, particularly ResNet-101, could be integrated into routine chest CT evaluations, providing added diagnostic value without requiring extra procedures or equipment.

By leveraging existing imaging infrastructure and reducing patient burden, this approach represents a meaningful advancement in breast cancer diagnostics.

 

The translation of the preceding English text in Chinese:

 

新研究探索利用深度学习通过常规CT扫描改善乳腺癌诊断

乳腺癌仍是全球女性中最常见的恶性肿瘤，也是癌症相关死亡的主要原因之一。根据美国癌症协会的数据，到2025年，美国将有近32万人被诊断为乳腺癌，超过4.2万人因此去世。早期发现对于提高生存率至关重要。然而，区分早期乳腺癌与良性乳腺肿块具有挑战性。此外，判断癌症是否已扩散至腋窝淋巴结（ALNM）对治疗方案制定和预后评估也至关重要。

传统的影像检查方法如乳腺X线摄影、超声（US）、磁共振成像（MRI）以及正电子发射断层扫描-计算机断层扫描（PET-CT）虽然常用，但存在局限性，包括患者不适、检查时间长、费用高，以及难以通过单一技术获得明确诊断。

相比之下，非增强胸部CT因常规体检和医院检查流程而被广泛使用于肺部、心脏和纵隔的评估。由于胸部CT通常包括整个乳腺区域，研究人员开始探索其在乳腺癌检测中的潜力，特别是结合人工智能（AI）技术。

研究目标：一个双重功能的深度学习模型

中国山东的孙静祥等研究团队开发了一个基于非增强胸部CT图像的深度学习模型，旨在预测乳腺肿块是良性还是恶性，并判断早期乳腺癌是否已扩散至腋窝淋巴结。

该回顾性研究分析了2018年至2023年间482例乳腺肿块患者的胸部CT图像，其中224例为良性，258例为恶性。在恶性组中，有91例已确认存在ALNM，167例未见转移。

所有患者均在手术前接受了非增强CT扫描。“这种方法无需使用造影剂，减少了对MRI或PET-CT等其他成像手段的依赖，节省了检查时间与成本，并最大限度降低了不必要的辐射暴露，”作者指出。

方法：应用ResNet深度学习架构

研究团队采用了三种常用于计算机视觉任务的残差网络（ResNet）架构：ResNet-34、ResNet-50和ResNet-101。所有模型均以乳腺区域图像为训练基础，经过尺寸调整与归一化处理以确保一致性。数据集被分为训练集（80%）、验证集（10%）和测试集（10%）。

每个模型需完成两项分类任务：

1. 区分良性与恶性乳腺肿块；

2. 预测早期乳腺癌患者是否存在ALNM。

模型性能通过准确率（ACC）、灵敏度（SEN）、特异度（SPE）、阳性预测值（PPV）、阴性预测值（NPV）以及ROC曲线下面积（AUC）进行评估。

结果：ResNet-101表现最佳

所有模型表现良好，但ResNet-101在两项任务中均表现优于其他模型。

在肿块分类方面：

ResNet-101的AUC为0.964，准确率为0.859，灵敏度为0.929，特异度为0.917。

ResNet-34和ResNet-50也表现出色，AUC均高于0.96。

在ALNM预测方面：

ResNet-101再次领先，AUC为0.951，准确率为0.874，灵敏度为0.865，特异度为0.909。

ResNet-50和ResNet-34的AUC分别为0.936和0.819。

研究发现，肿块分类的总体准确率略高于ALNM预测。作者认为，这可能是由于肿块的形状、边缘清晰度和密度等视觉特征更易识别，而淋巴结转移涉及更复杂的影像模式。

临床意义：迈向机会性乳腺癌筛查

该研究提出了一种实用方案：利用原本为其他健康评估而进行的非增强胸部CT扫描筛查乳腺异常。将深度学习模型整合进临床工作流程，可帮助放射科医生早期识别可疑乳腺病变或ALNM，无需额外成像检查或侵入性操作。

“非增强胸部CT具有多项优势，”作者指出。“无痛苦、不需要多种体位配合，空间分辨率高于超声。”

与增强CT或PET-CT相比，非增强CT辐射更低，避免了造影剂相关风险。相比乳腺MRI，其更具可及性与成本效益，尤其适用于资源有限的环境。

---

领域背景：AI与影像组学在乳腺成像中的应用

尽管已有多项研究利用增强CT结合影像组学或AI模型评估乳腺癌与ALNM，但鲜有研究聚焦于非增强CT或在同一模型中同时完成两项任务。此前基于影像组学的研究在ALNM预测中最高可达0.94的AUC，但需使用造影剂。

作者强调，他们的模型完全基于非增强CT，且能同时处理分类与转移预测，扩大了其临床适用性。

局限性与未来方向

尽管研究具有一定优势，但也存在一些局限性：

• 回顾性、单中心研究；

• 数据集中仅包括早期乳腺癌，且患者年龄分布不自然；

• 成像数据来自多个CT扫描仪，可能导致图像差异。

作者指出，未来需开展多中心、前瞻性研究以验证研究结果，并将继续扩展数据集、优化模型以提升ALNM预测能力。

结论：迈向可及的AI辅助诊断

本研究表明，基于非增强胸部CT图像的深度学习模型在识别乳腺癌及评估淋巴结转移方面具有准确性。特别是ResNet-101模型，可整合至常规胸部CT评估中，无需额外程序或设备，提升诊断价值。

通过利用现有成像基础设施、降低患者负担，该方法为乳腺癌诊断带来有意义的进展。

---

Reference:

Jingxiang Sun, Xiaoming Xi, Mengying Wang, Menghan Liu, Xiaodong Zhang, Haiyan Qiu, Youxin Zhang, Taian Fu, Yanan Du, Wanqing Ren, Dawei Wang, Guang Zhang

A deep learning model based on chest CT to predict benign and malignant breast masses and axillary lymph node metastasis. Biomol Biomed [Internet]. 2025 Mar. 17 [cited 2025 May 9];

Available from: https://www.bjbms.org/ojs/index.php/bjbms/article/view/12010

---

Important information:

We invite submissions for our upcoming thematic issues, including:

• Immune Prediction and Prognostic Biomarkers in Immuno-Oncology
• Artificial Intelligence and Machine Learning in disease diagnosis and treatment target identification
• Insulin for Type 1 and Type 2 Diabetes Mellitus: Pathophysiology, Complications and Treatment
• Cell-based Therapeutics

More news: Blog

Editor: Merima Hadžić