A Study on Predicting the Risk of Distant Metastasis in Breast Cancer Using AI-Generated Spatial Pathological Maps

A Study on Predicting the Risk of Distant Metastasis in Breast Cancer Using AI-Generated Spatial Pathological Maps

The goal of this observational study is to develop and validate an artificial intelligence (AI) model for predicting the risk of distant metastasis in patients with primary breast cancer. The main question it aims to answer is:

Can a multimodal AI model, trained on routinely available histopathological images, accurately predict the long-term risk of breast cancer metastasis?

Researchers will analyze existing hematoxylin and eosin (H&E) and immunohistochemistry (IHC) stained tissue slides from patients who underwent surgery between 2015 and 2025. Clinical data will be used to train the AI model and evaluate its performance in predicting metastasis.

Inclusion Criteria:

1. Female patients aged 18 years or older.
2. Histologically confirmed primary invasive breast carcinoma.
3. Underwent curative surgical resection (mastectomy or breast-conserving surgery) between January 2015 and December 2025.
4. Before initiating the neoadjuvant therapy, there was a retention of the primary tumor specimen.
5. Availability of high-quality, digitizable Hematoxylin and Eosin (H&E) stained whole-slide images (WSIs).
6. Availability of consecutive tissue sections from the same tumor block for multiplex immunohistochemistry (mIHC) staining (including markers such as Pan-CK, CD3, CD20).
7. Complete clinicopathological data and follow-up information must be available, including but not limited to: TNM stage, histological grade, molecular subtype (ER, PR, HER2 status), adjuvant treatment records, and clearly documented distant metastasis-free survival (DMFS) data.
8. A minimum follow-up of 5 years for patients with detailed information for distant metastasis events.

Exclusion Criteria:

1. Pure ductal carcinoma in situ (DCIS) without an invasive component.
2. Special histological subtypes of invasive carcinoma (e.g., metaplastic carcinoma) with distinct biological behaviors.
3. No original lesion samples were retained before neoadjuvant therapy.
4. Presence of contralateral breast cancer or a history of any other prior malignancy (except for cured non-melanoma skin cancer or carcinoma in situ of the cervix).
5. H&E or IHC slides with significant technical artifacts (e.g., fading, folds, heavy knife marks, tissue tearing, uneven staining) that preclude reliable image analysis.
6. Low tumor cellularity (e.g., tumor area < 10% in the scanned field of view).
7. Unavailable or unalignable consecutive tissue sections, preventing spatial registration of H&E and mIHC images.
8. Lack of essential clinicopathological or follow-up data required for model training or validation.