Mirai: An AI model to predict future breast cancer

6 min readAug 27, 2024

Breast cancer is one of the most common cancers among women worldwide, and millions of women undergo breast cancer screening every year to detect and treat the disease early. Mammography is currently the most commonly used screening method. However, existing breast cancer risk assessment models have certain limitations in accuracy and universality, which limits the effectiveness of screening strategies.

Limitations of existing risk assessment models

Lack of accuracy
Insufficient use of image information
Race and equipment differences

Mirai, a deep learning model developed jointly by MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) and the Jameel Clinic, solves the above problems.

Features of Mirai:

Multi-time point risk prediction

Mirai can predict breast cancer risk at multiple time points ( e.g., risk within 1, 2, 3, 4, and 5 years), thereby providing more comprehensive information for clinical decision-making.

Handling missing risk factor information

Mirai is able to process potentially missing risk factor information (such as age, family medical history, hormonal factors, etc.) and supplement model inputs by predicting these factors to ensure the accuracy of risk prediction.

Consistent risk assessment

Mirai uses a conditional adversarial training mechanism to ensure that it can perform consistent risk assessments on different types of mammography equipment. This means that no matter what device is used to take the mammogram, the model can provide relatively consistent risk prediction results.

Efficiently identify high-risk patients

Mirai performed well on multiple datasets, particularly in identifying patients at high risk of developing breast cancer within the next five years, and was more accurate than existing risk assessment models such as the Tyrer-Cuzick model.

Extensive international validation

Mirai demonstrated consistent performance across different ethnicities, ages, and breast density categories, demonstrating its potential application value in multiple clinical settings.

Technical methods

Dataset and model development

Dataset collection

Massachusetts General Hospital (MGH)
Karolinska University Hospital
Chang Gung Memorial Hospital (CGMH)

Model Architecture

Image Encoder: Each mammogram view is encoded using a shared ResNet-18 model.
Image Aggregator: Uses a Transformer model to aggregate the encoded information from different views into a comprehensive vector.
Risk Factor Prediction Module: Predicts traditional risk factors such as age, weight, and hormonal factors from mammograms.
Additive risk layer: combines the image aggregator output with risk factor information to predict a patient’s breast cancer risk over the next 5 years.

Model training and testing

Training process

The Mirai model is trained on the MGH dataset, using mammograms and corresponding risk factor information.
A conditional adversarial training mechanism is used to ensure consistent predictions of the model on different devices.

Model Evaluation

The models were tested on the MGH, Karolinska and CGMH datasets to evaluate their predictive accuracy (C index and area under the ROC curve).
Compare the performance of the Mirai model with other models (Tyrer-Cuzick model, Hybrid DL, and Image-Only DL).

Specific subgroup analysis

In the MGH dataset, subgroup analyses were performed by race (white, African American, and Asian American), age group, breast density category, and different devices to evaluate the consistent performance of the models.
In the Karolinska dataset, C-index calculation was performed by future cancer subtype (aggressiveness, HER2 status, etc.).

High-risk patient identification

High risk threshold setting

A 20% lifetime risk from the Tyrer-Cuzick model was used as the high-risk threshold.
The same specificity thresholds as the Tyrer-Cuzick model were set for the Image-Only DL, Hybrid DL, and Mirai models for sensitivity comparison.

High-risk patient identification performance

The sensitivity and specificity of each model in identifying high-risk patients were evaluated, and the performance of each model on different test sets (MGH, Karolinska, and CGMH) was compared.

Bias removal and feature importance analysis

Equipment deviation elimination

The conditional adversarial training mechanism is used to ensure the consistency of the model’s predictions on different mammography devices.
Evaluate the debiasing effect through the device identity classifier.

Feature Importance Analysis

Assess the importance of each risk factor in the Mirai model prediction and calculate the significance score of each risk factor.

Prospective research and model improvement directions

Prospective studies

The actual clinical application effect of the model needs to be further verified in large-scale clinical trials.

Model improvement direction

The model’s predictive accuracy was further improved by using 3D mammograms in conjunction with the patient’s imaging history.
Study how to adapt to mammography equipment from different manufacturers to ensure wide clinical application.

Experimental Results

Model performance

Overall Performance

The C-index of the Mirai model on three test sets (MGH, Karolinska, and CGMH) is 0.76, 0.81, and 0.79, respectively, showing higher performance than existing models such as Tyrer-Cuzick and Hybrid DL.
The area under the receiver operating characteristic curve (AUC) of the Mirai model was significantly higher than that of other models in identifying high-risk patients within 5 years.

Multi-time point risk prediction

The Mirai model performed well in predicting breast cancer risk within 1, 2, 3, 4, and 5 years, achieving high AUC values at each time point.
In the MGH dataset, the 5-year AUC of the Mirai model was 0.76, which was significantly higher than that of the Hybrid DL (P < 0.001) and Tyrer-Cuzick models (P < 0.001).

High-risk patient identification

Sensitivity and Specificity

In the MGH dataset, the Mirai model has significantly higher sensitivity than the Tyrer-Cuzick model and other deep learning models under the same specificity conditions. For example, the Mirai model identified 41.5% of high-risk patients within 5 years on the MGH test set, while the Tyrer-Cuzick model only identified 22.9%.
The sensitivity of the Mirai model is also significantly higher than that of the Image-Only DL model in the Karolinska and CGMH datasets. For example, in the CGMH dataset, the sensitivity of the Mirai model is 37.4%, while that of the Image-Only DL model is 24.5%.

Subgroup analysis

Race and Age Group

The Mirai model has consistent C-index performance across different races (white, African American, Asian American) and age groups (<40, 40–50, 50–60, 60–70, >70). For example, in the MGH dataset, the Mirai model has a C-index of 0.75 and 0.80 for white and Asian Americans, respectively, while the Tyrer-Cuzick model has a C-index of 0.64 and 0.54, respectively.

Breast Density Categories

The Mirai model performed consistently across different breast density categories (fatty, scattered fibroglandular, heterogeneously dense, and extremely dense), demonstrating its stability and reliability across different breast densities.

Future cancer subtypes

In the Karolinska dataset, the Mirai model’s C-index performance was consistent across different future cancer subtypes (aggressiveness, HER2 status, etc.), further validating its applicability across different cancer types.

……

For more info ↓