Similar Lesion Suggestion Engine

Purpose	Provide a type of 'reverse image' search for a highlighted lesion to provide the radiologist with similar appearing lesions and their pathologic outcomes.
Tag(s)
Panel	Breast Imaging
Define-AI ID	19060007
Originator	Hari Trivedi
Lead	Hari Trivedi
Panel Reviewers	Breast Panel
License	Creative Commons 4.0
Status	Public Comment
Organization	Emory University
RadElement Set	RDES144

Clinical Implementation

Value Proposition

Many rural and international radiology practices suffer from a lack of breast-trained radiologists. Patients undergoing breast cancer screening at these sites may be under- or over-diagnosed dependent upon the skill level of the interpreting radiologists. Because over 90% of studies are negative on screening mammography and at least 75% of studies are negative on diagnostic imaging, non-breast radiologists may not have the breadth of experience required to identify nuanced differences that separate malignant from benign lesions. A tool that enables radiologists to specify a lesion of interest and receive a list of similar-appearing lesions with biopsy-proven pathologic outcomes would increase confidence in management decisions and curb unnecessary recall imaging and biopsies.

Narrative(s)

A 45 year-old woman arrives for annual screening mammography in a rural area. The radiologist reviewing the case notices calcifications in the left breast but has not encountered this specific appearance before and is leaning towards bringing back the patient for recall imaging. The radiologist highlights the area of concern on the mammogram and a DL algorithm provides a list of similar appearing lesions which the radiologist is able to review. All of the similar appearing lesions were benign on pathology and the radiologist therefore determines that this lesion is of no concern. The patient returns to annual screening mammography and an unnecessary recall study is avoided. 

Workflow Description

A radiologist is viewing a screening mammogram in their standard PACS software. A potential abnormality is identified and the radiologist draws a bounding box around the lesion. The bounding box can be of any size or shape specified by the radiologist. This image is then processed by either a local or cloud DL algorithm and an internal database is reviewed for similar lesions. A list of similar appearing lesions is then displayed in a table format with a ‘similarity score’ and final pathologic diagnosis for each lesion. If no similar lesions are found, the algorithm will return no results. The radiologist will use this information to help guide their decision and can close the panel to continue with their regular workflow.

Considerations for Dataset Development

Sex	Female
Age	> 18 years old
Pathologic diagnosis	varied
View(s)	Full-field diagnostic views such as CC, MLO, LM, XCCM, XCCL, cleavage view, etc. Clinical Note: Spot and magnification views should be identified separately as these have higher resolution, smaller fields of coverage, and higher incidences of malignancy which would bias the algorithm
Bounding box dimensions over lesion of interest	varied

Technical Specifications

Inputs

 

DICOM Study

Procedure	Diagnostic Mammography
Views	CC, MLO, LM, XCCM, XCCL, Cleavage view, etc.
Data Type	DICOM
Modality	MAMMO
Body Region	Chest
Anatomic Focus	Breast

Secondary Inputs

 

Bounding Box Around a Lesion

Definition	Clinician selected bounding box around a lesion of interest
Data Type	DICOM
Value Set	N/A
Units	N/A

Primary Outputs

 

Similar Lesions in the Database

 

 

 

Similarity Score

 

 

 

 

Pathologic Diagnosis For Similar Lesions

 

 

 

Secondary Outputs

 

Patient Age

 

 

 

 

Breast Cancer History

 

 

 

Genetic Risk Factors for Breast Cancer