Mammograms project

These slides: bit.ly/hdl-mammograms

Health Data Lab: hdl.cs.uit.no/

Lars Ailo Bongo, �Professor in Health Technology, �Department of Computer Science

​

in collaboration with: Mike Voets, Kajsa Møllersen, �Kvinner og Kreft, Kreftregisteret

​

UNN AI meeting, 10.03.20

Hypothesis: two bad tests makes a good test

Our goal was to reduce mammography screening recall

Pilot project: how well does state-of-the-art machine learning methods works on Norwegian mammograms?

We need Norwegian Mammograms�(initial naive plan)

UNN

UiT

Mammograms

...16 months later...

Anders Høydalsvik: Helse Nord IKT

Solveig Hofvind:

Mammografi-�programmet

Linda S. Hansen:�BDS UNN

With great help from:

​

​

​

​

​

​

​

​

​

​

​

​

​

...and many others

The process to get anonymized data

1. Create a project team that includes a facilitator at UNN.
2. Make a plan for what data to retrieve (data transfer agreement).
3. Get “not their concern confirmation” from REK.
4. Get samtykke. Registers with existing approvals makes this much easier.
5. Get approval from UNN PVO. Asking for all of the data is unusual.
6. Apply to get data, including examination IDs and journal data, from a register.
7. Make a service order. Only UNN can do this.
8. Use examination IDs to retrieve images from PACS.
9. Link images with journal data (from register).
10. Anonymize images and journal data by removing fields.
11. Copy anonymized data to USB disk.

We have a pilot dataset, but need even more data for deep learning

• 350.000 mammography images �= 87.500 examinations
• But only:
• 8750 (or less) recalls
• 437 cancer cases
• even fewer for specific breast cancer subtypes

​

Use of deep learning and big data in the Norwegian breast cancer screening program

Drammen

Ålesund

Hamar

Lillehammer

UNN

Østfold Kanes

Kristiansand

St. Olavs

​

Collect mammograms from most BDS’ in Norway (not anonymized)

St. Olavs

Mammograms

Hemit

800.000

Mammograms

Kreftregisteret

2.000.000

Mammograms

UNN

OuS

Helse Nord IKT

Sykehuspartner

10x more costly than budgeted �(and much more time consuming)

Initial evaluation using UNN pilot data: Pre-trained model works for Norwegian mammograms

While waiting for data we attempted to replicate well known deep learning paper

In ongoing research we are counting cells in UNN pathology (HE) slides using pre-trained models

Project received Helse Nord funding (PI: Thomas Kilvær).

with: Nikita Shvetsov, Edvard Pedersen, �Thomas Kilvær, Lill-Tove Rasmussen Busund

Summary

• The Health Data Lab:
• Integrate omics data with medical images
• Data management
• Apply existing methods and models for new analyses
• Getting access to UNN data should be easier
• Make guidelines and routines
• Better technical solutions (Helse Nord IKT DMA?)

​

• Lab homepage: http://hdl.cs.uit.no/
• My contact information: Lars Ailo Bongo <lars.ailo.bongo@uit.no>