Skip to main content

Deep learning for mammography: Improving accuracy and productivity in breast cancer diagnosis.

Periodic Reporting for period 1 - MAMMO1 (Deep learning for mammography: Improving accuracy and productivity in breast cancer diagnosis.)

Reporting period: 2019-01-01 to 2019-04-30

Mia (previously Mammo1) is deep-learning-based software designed to help radiologists and breast screening programmes. It uses convolutional neural networks to scan Full Field Digital Mammograms (FFDM), to identify patterns indicating cancer in a case-wise manner, offering a recall suggestion and highlighting the areas on which the decision is based. It can serve as a second or third reader, in triaging (making priority lists by detecting the most potentially malignant cases) and as a quality assessment tool.
The Phase I H2020 SME Instrument helped us to assess technical, commercial and financial aspects to ensure market uptake; study IP protection, research our freedom-to-operate; and identify potential risks we may encounter during development, commercialisation and system scale-up, and design mitigation measures.
Mia’s prototype has already been tested in a retrospective clinical study, validated externally. This demonstrated sensitivity and specificity of 90% and 89% respectively. These results surpass those of the best known competitors and led to successful awarding of the CE Mark class IIa for Medical Devices, and winning ‘Best New Radiology Software’ as awarded by the radiology community Aunt Minnie.
Mia can achieve robust generalisability across all major mammography hardware vendors and global screening settings.
We have investigated the requirements to integrate our software into hospital workflows and designed two deployment architectures, on-site and cloud-based solutions, to cover all options while ensuring security requirements are met.
To assess Mia’s generalisability, integration and accuracy we aim to perform prospective clinical trials. Several European hospitals are already interested in participating.
Mia’s performance surpasses previous Computer-Assisted Diagnosis (CAD) solutions and the average single-reading radiologist. Designed to act as a second reader in blinded double-read settings with arbitration, reduces up to 50% in workload and manpower costs, while maintaining the low recall rates expected in population-based screening programmes. Additionally, Mia can help prioritise potentially malignant cases prior to review, improving speed of diagnosis. The wider implication is that nation-wide screening programmes suffering from the current breast screening workforce crisis will have the opportunity to provide high quality, fast and accurate screening for any woman, anywhere, with reduced geographic variability in outcomes.
How Mia Works
Hospital deployment architectures for Mia
Mia's performance compared to competitors and single radiologists