AI can improve the detection and the prevention of breast cancer

Prevention and control of breast cancer

Breast cancer is the most common cancer among women in both developed and developing countries. More than 508,000 women worldwide died from breast cancer in 2011 (World Health Estimates, OMS 2013). One of the keys to the fight against breast cancer is prevention through mammography.

AI researchers are using advances in convolutional neural networks to classify patients with and without cancer from mammograms.

Breast cancer: prevention and control

Explainable Deep Learning in Breast Cancer Prediction

Detect tumors better than experts

Since 2018, Google DeepMind, DeepHeath, NYU, Baidu, Stanford have published research papers on CNN that allow to determine from mammograms, the presence or not of a breast cancer tumor.

Google’s AI model surpassed the six human experts. DeepHealth’s model surpassed the five breast imaging experts. The New York University model outperformed 12 treating radiologists with 2–25 years experience, one resident and one medical student.

Researchers at Google, DeepHealth and NYU all emphasized that their systems are not intended to replace radiologists but rather to support them in interpreting breast cancer screening exams.

Google, NYU, DeepHealth competition

Three Papers in the Eye of the ‘AI Breast Cancer Detection’ Storm

NYU open-source model

NYU open-sources breast cancer screening model trained on over 200,000 mammography exams

DeepMind model

International evaluation of an artificial intelligence system to identify breast cancer in…

Baidu reasearch

Baidu Research’s breast cancer detection algorithm outperforms human pathologists

Stanford research

Close calls: How AI could improve breast tumor diagnosis — Scope

Predict breast cancer risk up to 5 years before diagnosis

Regina Barzilay , a researcher and professor at MIT, has also used a CNN to detect whether or not a patient has breast cancer, but her team has gone one step further by using a CNN to predict breast cancer risk up to 5 years before diagnosis.

Using AI to predict breast cancer and personalize care

Model 2 is Model 1 augmented with risk factors

The models developed by the MIT teams outperform the reference model used by American therapists, the “Tyrer-Cuzick” or IBIS tool. This Tyrer-Cuzick model allows the calculation of the probability that a woman will develop breast cancer in 10 years and during her lifetime.

Ikonopedia IBIS: Online Tyrer-Cuzick Model Breast Cancer Risk Evaluation Tool

• Model 2 perform better than Tyrer-Cuzick model

• Model 2 is more generalizable than Tyrer-Cuzick for african american women

How to overcome the limitations of AI?

For Régina Barzilay, AI has super powers, but researchers face three problems that will be the focus of research in the years to come.

Problem 1: not enough data

=> the solution is to perform fine tuning.

Fine-Tuning ResNet for Breast Cancer Classification from Mammography

• Problem 2: difficulty to generalize with new data

=> the solution is to use generative adversarial network (GAN)

Breast cancer detection using synthetic mammograms from generative adversarial networks in…

• Problem n°3 : deep learning models are black boxes

=> the solution is to use “interpretable neural network” (Grad Cam) to know what neural networks see in order to perform their classification.

[PDF] Visual interpretability for patch-based classification of breast cancer histopathology images…

Article written from this video : Regina Barzilay, power and limits of Machine Learning

To learn more about Régina Barzilay, Regina Barzilay: Deep Learning for Cancer Diagnosis and Treatment | Lex Fridman Podcast #40