AsymMirai: Interpretable Mammography-based Deep Learning Model for 1-5-year Breast Cancer Risk Prediction

Jon Donnelly ¹, Luke Moffett ¹, Alina Jade Barnett ¹

⁰From the Departments of Computer Science (J.D., L.M., A.J.B., C.R.) and Electrical and Computer Engineering (C.R.), Duke University, 308 Research Dr, LSRC Building D101, Duke Box 90129, Durham, NC 27708; Department of Radiology and Imaging Services, Emory University, Atlanta, Ga (H.T.); Department of Radiology, Harvard University, Cambridge, Mass (F.S.); and Department of Radiology, Duke University School of Medicine, Durham, NC (J.L.).

Radiology +

|

March 19, 2024

View abstract on PubMed

Summary

A new deep learning model, AsymMirai, uses breast tissue asymmetry to predict cancer risk, achieving performance similar to the complex Mirai algorithm. This interpretable approach enhances breast cancer risk assessment by identifying key imaging markers.

Area Of Science

Radiology and Medical Imaging
Artificial Intelligence in Healthcare
Oncology

Background

Mirai, a deep learning algorithm, excels at predicting short-term breast cancer risk but functions as a black box, potentially leading to overreliance and misdiagnosis.
The interpretability of deep learning models in medical imaging is crucial for clinical trust and accurate diagnostic decision-making.

Purpose Of The Study

To investigate if bilateral dissimilarity (differences between left and right breast tissue) is a key factor in Mirai's risk predictions.
To develop AsymMirai, a simplified, interpretable deep learning model based on bilateral dissimilarity for breast cancer risk prediction.
To evaluate AsymMirai's performance in approximating Mirai's 1-5 year breast cancer risk prediction accuracy.

Main Methods

Retrospective analysis of 210,067 screening mammograms from 81,824 patients in the EMory BrEast imaging Dataset (EMBED).
Development of AsymMirai, incorporating local bilateral dissimilarity as an interpretable module, to predict 1-5 year breast cancer risk.
Comparison of AsymMirai and Mirai risk scores using Pearson correlation and Area Under the Receiver Operating Characteristic Curve (AUC) analysis.

Main Results

AsymMirai demonstrated strong correlation with Mirai's risk scores (1-year: r=0.6832, 4-5 year: r=0.6988), indicating similar reasoning.
AsymMirai achieved comparable performance to Mirai: 1-year AUC 0.79 (Mirai 0.84) and 5-year AUC 0.66 (Mirai 0.71).
In a subgroup with consistent AsymMirai reasoning, a 3-year AUC of 0.92 was achieved, highlighting potential for high accuracy in specific cases.

Conclusions

Localized bilateral dissimilarity is a key imaging marker that approximates Mirai's predictive power and underlies its reasoning process.
AsymMirai offers an interpretable alternative for breast cancer risk prediction, demonstrating that simplified models can achieve comparable performance to complex black-box algorithms.
The findings support the clinical utility of interpretable AI in enhancing breast cancer risk assessment and potentially improving diagnostic accuracy.