- About
- Our solutions
- Resource Center
April 19 2022
|
Reading time 9 minutes
Contents
Breast cancer is the most widespread type of cancer faced by women1. Regular screening with visualization methods including mammography allows reducing mortality2.
However, overdiagnosis is an issue: breast tumors can be considered as cancer by mistake. They are not life-threatening and would not be found without screening3.
Artificial intelligence (AI) helps radiologists to solve relevant issues in mammography. Doctors use advanced technologies for interpretation of images and assessment of breast cancer risk, increasing accuracy of diagnosis and prognosis.
To teach artificial intelligence to identify problems in breast images, specialists mark the images in advance. They mark problem zones manually and indicate a diagnosis. The algorithm will further use these marks.
What diagnostic tasks in mammography does artificial intelligence tackle4?
Lump identification and classification. Algorithms help specify the form, size and edges of suspicious areas in breasts. The neural network carries out segmentation, dividing images into parts with healthy tissues and problems.
Microcalcification identification. Microcalcifications are small deposits of calcium salts in breast tissues. In CT images they look like white spots. Timely identification of microcalcifications is important in mammography because they can be a sign of a malignant process. AI detects extremely small formations which the human eye cannot see: from 0.1 to 1 mm. Algorithms analyze the number of microcalcifications and zones of their accumulation. This increases accuracy of a radiologist’s work.
Breast density assessment. High breast density is a cancer risk factor. AI assesses mammographic density and indicates percentage of abnormal tissue. The analysis is made in different projections.
Predicting the risk of cancer. Artificial intelligence uses mammography results, clinical data and risk factors of a patient to assess probability of cancer. This helps doctors to correct treatment.
Programs based on machine learning continuously process big volumes of data. Every new mammogram is another step towards more accurate decisions for AI.
Scientists from Iran developed artificial intelligence which helps a radiologist to correctly assess breast cancer risk. The neural network analyzes mammography results and clinical variables, such as age and family health history5.
AI assessed the data of 655 women, 196 of which had cancer, and 459 women had benign tumors. Biopsy was used to verify the diagnosis5.
The neural network was able to differentiate between benign tumors and cancer with high accuracy. AUC was used for assessment. This indicator estimates accuracy of classification of true positive (patient with breast cancer) and false positive (healthy patients with diagnosed cancer) results5.
AUC shows probability that a random heathy patient is more rarely diagnosed with cancer than a random patient who really has it. The higher AUC is, the more accurately the algorithm differentiates between a normal condition and a problem. The indicator was found to be 0.9555.
Using AI as an additional instrument, a radiologist can correctly assess a risk of problems and diagnose with higher confidence5.
There are certain difficulties with organizing screening in mastology. A radiologist spends time mainly on diagnosis of healthy patients. At the same time, women with the first signs of problems which are often not very noticeable can get a false negative result6.
Swedish scientists made artificial intelligence their assistant in triaging and published results in Lancet Digital Health. The researchers examined 7,364 women aged between 40 and 70, 547 of which had cancer, and 6,817 healthy women were included in the control group6.
The developed AI sorts mammograms in several stages6:
During initial analysis, the algorithm assesses breast cancer probability. The assessment is a number from 0 to 1, where 1 is the highest level.
AI recognizes images with low probability of cancer as normal and “deletes” them from processing, placing them to a group without a radiologist.
Mammograms with high risk are analyzed twice with involvement of two radiologists.
If a suspicious formation is identified, specialists discuss which category should be assigned to the image: whether it is normal or more detailed examination is necessary.
To verify the diagnosis, a patient undergoes an MRI scan.
The algorithm more than halved the time for establishing a diagnosis, excluding normal mammograms at initial stages. For 60%, 70% and 80% of images with low probability of problems included in the group without a radiologist, the share of unidentified problems was 0%, 0% and 3% respectively6.
AI made it possible not to use the expensive MRI technology as the “first line” method and use it only for eliminating false negative results6.
An open-science crowdsourced Digital Mammography DREAM Challenge platform set a relevant task for participants: to offer an algorithm which will allow improving breast cancer screening accuracy7.
The contest was organized by technology organizations such as Sage Bionetworks and IBM Research, Kaiser Permanente Washington Health Research Institute, and the Icahn School of Medicine at Mount Sinai8.
Digital Mammography DREAM Challenge took place in several stages:
Over 1,150 scientists and developers took part in the competition stage. Kaiser Permanente Washington Health Research Institute provided access to the database of 640,000 anonymized digital mammograms of 86,000 women. Participants taught their algorithms based on these data and submitted them for assessment7.
At the final stage of independent algorithm verification, data of women who underwent screening in Stockholm region of Sweden were used. The images contained information on presence or absence of cancer, but tumor location was not specified9.
The team of the French company Therapixel became the winner. The team was headed by senior fellow Yaroslav Nikulin. It managed to develop the algorithm with the accuracy of 80.3% and 80.4% for the two types respectively7.
Women in whose breasts neoplasms were found may need surgery. The decision may be made in uncertainty because there is a risk that the tumor is malignant.
Scientists from the North Cochrane Center and the Norwegian Institute of Public Health decided to find out how widely overdiagnosis of breast cancer is spread in Denmark. Results published in Annals of Internal Medicine show the following: One in three invasive tumors is not clinically relevant (overdiagnosis totaled 48.3%)10.
Artificial intelligence helps identify formations with a high risk of being malignant at earlier stages, and avoid unnecessary and painful surgeries. Employees of Harvard Medical School and Massachusetts Institute of Technology developed an algorithm which predicts probability of malignization11.
AI uses different diagnostic information for analysis11:
clinical data: height, weight, age, personal and family health history;
mammography results: neoplasm mass, asymmetry, calcifications;
biopsy results;
surgical pathology reports.
All data were verified by a radiologist. AI assigned high or low risk of malignization to each case of affected breasts11.
Scientists found out that patients with low risk could be under supervision, and other patients needed a surgery. In this case 97.4% of neoplasms would be correctly diagnosed as cancer during the surgery. In case of a benign tumor, 30.6% of surgeries could be avoided11.
A research published in Journal of Breast Imaging is concerned with radiologist burnout. The Society of Breast Imaging (SBI) suggested that 370 doctors take part in survey to assess their emotional state at workplaces12.
Breast radiologists aged between 36 and 45 had the highest burnout level. Burnout was identified in 83.2% of respondents by at least one parameter12.
AI-based Mammography is a service that effectively performs breast imaging tasks. Smart digital product tools by SberMedAI create a comfortable working environment for health professionals:
making it easier to see problems by outlining suspicious areas in a mammogram;
carrying out automatic neoplasm assessment based on BI-RADS scale;
providing assistance in prognosis, determining cancer probability;
giving recommendations for further examinations;
reducing the time for test result analysis to one minute or less;
providing informational support for doctors, helping them to make right decisions irrespective of experience.
Automation of routine processes with Mammography service helps reduce workload on breast radiologists. Doctors have less stress and make sensible decisions due to assistance in diagnosis.
Sources
Bernhard Wörmann. Breast Cancer: Basics, Screening, Diagnostics and Treatment. Med Monatsschr Pharm. 2017 Feb; 40(2):55-64. English, German. PMID: 29952495.
Paula B Gordon. Breast Cancer Screening. Can Fam Physician. 2019 Jul;65(7): 457-459. PMID: 31300424; PMCID: PMC6738458.
Ismail Jatoi, Paul F Pinsky. Breast Cancer Screening Trials: Endpoints and Overdiagnosis. J Natl Cancer Inst. 2021 Sep 4;113(9):1131-1135. doi: 10.1093/jnci/djaa140. PMID: 32898241; PMCID: PMC8633447.
Yu-Meng Lei, Miao Yin, Mei-Hui Yu et al. Artificial Intelligence in Medical Imaging of the Breast. Front Oncol. 2021 Jul 22;11:600557. doi: 10.3389/fonc.2021.600557. PMID: 34367938; PMCID: PMC8339920.
Mojtaba Sepandi, Maryam Taghdir, Abbas Rezaianzadeh, Salar Rahimikazerooni. Assessing Breast Cancer Risk with an Artificial Neural Network. Asian Pac J Cancer Prev. 2018 Apr 25;19(4): 1017-1019. doi: 10.22034/APJCP.2018.19.4.1017. PMID: 29693975; PMCID: PMC6031801.
Karin Dembrower, Erik Wåhlin, Yue Liu et al. Effect of Artificial Intelligence-based Triaging of Breast Cancer Screening Mammograms on Cancer Detection and Radiologist Workload: a Retrospective Simulation Study. Lancet Digital Health. 2020 Sep;2(9):e468-e474. doi: 10.1016/S2589-7500(20)30185-0. PMID: 33328114.
DREAM Challenge Results: Can Machine Learning Help Improve Accuracy in Breast Cancer Screening? IBM [Online]. URL: https://www.ibm.com/blogs/research/2017/06/dream-challenge-results/ (request date: 02/03/2022).
The Digital Mammography DREAM Challenge: Synapse [Online]. URL: https://www.synapse.org/#!Synapse:syn4224222/wiki/401743 (request date: 02/03/2022).
Thomas Schaffter, PhD; Diana S. M. Buist, PhD, MPH; Christoph I. Lee, MD, MS; et al. Evaluation of Combined Artificial Intelligence and Radiologist Assessment to Interpret Screening Mammograms. JAMA Netw Open. 020;3(3):e200265. doi:10.1001/jamanetworkopen.2020.0265
Karsten Juhl Jørgensen, Peter C Gøtzsche, Mette Kalager, Per-Henrik Zahl Breast Cancer Screening in Denmark: A Cohort Study of Tumor Size and Overdiagnosis. Ann Intern Med. 2017 7 March; 166 (5): 313-323. doi: 10.7326/M16-0270. Epub 2017, 10 January. PMID: 28114661.
Manisha Bahl , Regina Barzilay, Adam B. Yedidia et al. High-Risk Breast Lesions: A Machine Learning Model to Predict Pathologic Upgrade and Reduce Unnecessary Surgical Excision. Radiology. 2018 Mar; 286(3):810-818. doi: 10.1148/radiol.2017170549. Epub 2017, 17 October. PMID: 29039725.
Alison L Chetlen, Tiffany L Chan, David H Ballard et al. Addressing Burnout in Radiologists. Acad Radiol. 2019 Apr; 26 (4): 526-533. doi: 10.1016/j.acra.2018.07.001. Epub 2018, 31 July. PMID: 30711406; PMCID: PMC6530597.
