Breast Lesion Malignancy Evaluation

Purpose

To automate classification of breast lesions into probability/likelihood of malignancy at time of screening mammography, US

Tag(s)


Panel

Breast Imaging

Define-AI ID

19060004

Originator

Syam Reddy

Panel Chair

Beth Burnside

Panel Reviewers

Breast Imaging Panel

License

Creative Commons 4.0

Status

Public Comment

 
 

 

Clinical Implementation


Value Proposition

Mammography, sonographic and MRI lesions may have many varying appearances that can be difficult to distinguish between benign and malignant. AI with deep learning can help differentiate lesions with low risk from high risk types or possibly predict cancer type and thereby more consistently help drive which patients require further diagnostic evaluation and ultimate biopsy. We can also add best practice recommendations to these lesions of concern. 

Narrative(s)

45 yr old high risk female presents for screening ultrasound. On her screening, a hypo echoic mass with partially circumscribed borders is seen. There is no vascularity and no posterior acoustic enhancement. AI views images or area of concern and give a percentage for a concerning lesion that would require further diagnostic evaluation. Based on lesion appearance and through deep learning, AI would give recommendation of 6 month follow up US as a best practice recommendation.

Workflow Description

Screening mammogram, US or MRI lesion images are sent to AI after selected as a possible abnormality by reading radiologist. AI reviews using prior deep learning experience for lesions analysis. A probability /likelihood percentage is given for each questioned lesions. Best practice guidance can be given for specific lesion type which would result in more appropriate diagnostic work up for patients.

Considerations for Dataset Development


Procedures(s): Screening mammography to include screening ultrasound and Breast MRI

Indication: Breast cancer screening and early detection


Breast anatomy: {No prior surgery, no implants}


Age: 40 years and older


Views: {Standard craniocaudal, Standard mediolateral}

Technical Specifications


Inputs

DICOM Study

Procedure

breast ultrasound,

Views

Standard craniocaudal and mediolateral, long/short or radial/antiradial US views,

Data Type

DICOM

Modality

US

Body Region

Chest

Anatomic Focus

Breast

Pharmaceutical

N/A

Scenario

N/A



Primary Outputs

Suspicious Breast Lesion Detection

RadElement ID

Definition

Identify suspicious lesions in screening situations that require further evaluation

Data Type

Categorical

Value Set

0-Benign lesion

1-Malignant lesion

2-Indeterminate

Units

N/A


Secondary Outputs



Breast Lesion Classification

RadElement ID

Definition

Lesion evaluation with radiology and pathology data

Data Type

Categorical

Value Set

1-Concordant benign

2- Concordant but high risk

3-Discordant

Units

N/A


Follow Up Plan Recommendation

RadElement ID

Definition

Recommend follow up action based on the breast lesion finding

Data Type

Categorical

Value Set

0- Unknown

1- 6 month follow up

2- Routine annual

3- 6 month follow up with possible breast MRI

4- Surgical excision

5- Repeat biopsy

Units

N/A

Future Development Ideas


Develop algorithm for those lesion biopsied to compare with pathology result data pulled from EMR and given recommended for concordance or treatment planning. Use lesion analysis to compare with pathology results for concordance.