BCR-Net | A New Model to Predict Breast Cancer Recurrence 

A research group at the Wake Forest School of Medicine (NC, U.S.) has developed an AI model that predicts the likelihood of breast cancer recurrence. This model offers a cost-effective and rapid method for pathologists to assess the best treatment options for breast cancer patients. 

Breast Cancer: Past and Present Perspectives 

Breast cancer remains a global health challenge, causing over 670,000 deaths in 2022, ranking it the second leading cause of cancer-related mortality worldwide. The incidence of breast cancer has been slowly on the rise, with about a 1% increase per year in the number of cases between 2012 and 2021.

To address this growing health burden, the WHO has launched an initiative aimed at reducing breast cancer mortality by 2.5% annually, intending to save 2.5 million lives between 2020 and 2040 across the globe. Central to this initiative is personalized medicine, where treatment plans are tailored to diverse patient groups with distinct genetic or clinical profiles.  

Patients with breast cancer are categorized into four subtypes that each respond differently to available treatments. 

For example, hormone receptor-positive breast cancers, which is the most common subtype, are typically treated with hormonal therapies such as tamoxifen, letrozole or aromatase inhibitors.

However, other than cancer subtypes, there is another critical element that needs to be considered here: determining whether a patient has a high risk of recurrence.  

Chemotherapy is a standard treatment for all cancer, however, since this form of therapy comes with unpleasant side effects, it is generally avoided for patients with a low risk of recurrence. For high-risk patients, whose cancer is likely to come back, patients often receive chemotherapy alongside other targeted treatments.   

Traditionally, recurrence risk is analyzed using a genetic assay known as Oncotype DX (ODX). This test identifies the overexpression of specific genes associated with cellular growth and recurrence risk. However, despite its efficacy, ODX is expensive and time-consuming, limiting its accessibility, especially in low-resource settings.  

AI has been proposed as a way to overcome these limitations.  

AI Could Help Predict Cancer Recurrence

AI presents a transformative opportunity in breast cancer diagnostics. Multiple research groups have investigated the use of AI to automate the analysis of histopathological slides, examining and analyzing cellular features to assess their association with recurrence risk. 

By training deep learning models via supervised learning on annotated datasets of whole-slide images (WSIs), researchers can predict recurrence risks efficiently. However, this requires exhaustive manual labelling of WSIs, which is incredibly time-consuming.  

To overcome this challenge, researchers recently introduced Breast Cancer Recurrence (BCR)-Net, a weakly supervised deep learning framework, in a study published in PLoS One. 

BCR-Net leverages vast amounts of partially labelled datasets rather than relying on a smaller amount of fully annotated data. This approach (weakly supervised learning) prioritizes the analysis of broader patterns over precise details, reducing the time and effort required for data preparation while maintaining high predictive accuracy. 

So, How Does BCR-Net Work?  

BCR-Net operates by analyzing histopathological WSIs, identifying regions or patches contributing to meaningful prediction scores, and thus assessing risk. Simultaneously, in its risk assessment, the model disregards non-discriminative regions or patches that do not contribute to predictions. 

The authors of the paper emphasize that BCR-Net’s innovative ability to identify discriminative patches from WSIs autonomously distinguishes it from other AI models in the field. This is achieved through a type of deep neural network framework called convolutional neural network (CNN)-scorer, which incorporates multiple ‘hidden’ layers to perform linear transformations and non-linear activation functions to the data at each step.  

The step by step: 

1. Linear Transformations: BCR-Net applies linear transformations to extract basic features from discriminative patches of the WSIs.  
2. Non-Linear Activation: Non-linear activation functions are used to emphasize significant features, suppressing noise from the non-discriminative patches. This step ensures that only the most relevant data is highlighted. 
3. Multiple Instance Learning: After significant feature extraction, the CNN-scorer employs multiple-instance learning to classify the WSIs. This is done by pooling analyses from individual patches to determine the overall risk category.

Attention-based pooling is a unique component of BCR-Net. The model assigns attention-based weights to patches based on their importance for risk prediction. The model then adaptively learns how to combine these instances into a single output, rather than using pre-defined calculations, ensuring that the model flexibly integrates the most relevant features for accurate predictions. 

Crucially, the attention weights make the model interpretable, providing clear insights into how the AI arrives to its decisions. This distinct feature sets BCR-Net apart from other CNNs, which are often susceptible to the “black box” phenomenon. Unlike models such as CLAM, which were trained using features from common objects irrelevant to cells, BCR-Net uniquely focuses on domain-specific features, enhancing its reliability and precision. 

Dissecting the Statistics

BCR-Net was trained on 14,000 differently sized patches extracted from 70 WSIs, and its accuracy in scoring recurrence risk was evaluated and compared against two state-of-the-art AIs: CLAM and TransMIL.   

Using 448×448 and 224×224 square patches, the group achieved remarkable accuracy, with an average AUC of 76%. Additionally, BCR-Net accurately classified WSIs as either low-risk or high-risk in 80% of cases, outperforming its comparator models. 

Transforming the Future of Breast Cancer Treatment 

The results obtained with BCR-Net pave the way for future AI tools in cancer recurrence risk prediction. It achieves state-of-the-art accuracy, whilst mitigating the need for labor-intensive genomic assays or fully annotated datasets for supervised machine models. This could cut back time spent on diagnostic procedures, enabling rapid treatment delivery to patients who require urgent attention.   

Leveraging technologies like BCR-Net can enable pathologists to quickly and effectively identify patients who would benefit from chemotherapy-based regimens, whilst avoiding unnecessarily aggressive treatments for those with low recurrence risks.  

The adoption of such AI tools has the potential to enhance the quality of life for breast cancer patients and transform the future of personalized cancer care.