Skip Navigation Bar

NLM Announces Pill Image Recognition Request for Information (RFI)

Responses, to Help with Project Planning Purposes, Due May 15, 2015

As part of its ongoing work to develop an infrastructure and tools for identifying prescription pills, the National Library of Medicine announces a new Request for Information (RFI), the details of which appear below. This is NOT a solicitation for proposals, proposal abstracts, or quotations. The purpose of this RFI is to obtain knowledge and information for project planning purposes. The government will not award a contract on the basis of this notice, or otherwise pay for information solicited by it. Proprietary information should be clearly marked. The requested information is for planning and market research purposes only and will not be publicly released.

This Request for Information (RFI) is a pilot for a forthcoming Pill Image Recognition Challenge for visually identifying pills. The Challenge has as its main objective the development and discovery of high-quality software that matches images of unknown prescription pills to images in the RxIMAGE database. The pilot will help ensure that the Challenge is successful.

PILL IMAGE RECOGNITION – REQUEST FOR INFORMATION (RFI)

Unidentified and misidentified prescription pills present challenges for patients and professionals. Unidentified pills can be found by family members, health professionals, educators, and law enforcement. The nine out of 10 US citizens over age 65 who take more than one prescription pill can be prone to misidentifying those pills. Taking such pills can result in adverse drug events that affect health or cause death.

To reduce such errors, any person should easily be able to be confirm that a prescription pill or a refill is correct. For example, a person should be able to easily verify–or not–that a refill that has a different color, shape, or size is just a different generic version of the same medication he or she was already taking.

To help address these problems, the National Library of Medicine (NLM) Computational Photography Project for Pill Identification (C3PI) is developing infrastructure and tools for identifying prescription pills. The infrastructure includes NLM’s RxIMAGE database of freely available, high quality prescription pill images and associated pill data. One tool is the freely accessible RxIMAGE API (Application Programming Interface) for text-based search and retrieval of images and data from the RxIMAGE database.

NLM now seeks to expand the toolset to include smart phone apps to visually search for and retrieve pill images and data. A person will photograph an unknown prescription pill, possibly under poor lighting conditions, from an angle, or at low resolution. The app will return one or more RxIMAGE images and data that are most likely to match the photographed pill.

Example:

(a) (a) User takes photo of front of unknown prescription pill (b) (b) Smart phone app returns one or more possible matches – in this case the images of the fronts and backs of four pills

(a) User takes photo of front of unknown prescription pill
(b) Smart phone app returns one or more possible matches – in this case the images of the fronts and backs of four pills

This RFI is extended to private industry, academia, national laboratories, and government agencies. Respondents to the RFI are asked to submit executable software that matches consumer images – photos of pills taken by cell phone digital cameras – with reference images obtained from images in the RxIMAGE database. It is anticipated that respondents will include professionals and students, individually or in teams, in computer vision and computer graphics working on content-based image retrieval. Respondents to this RFI may also participate in the Challenge.

The RFI and the Challenge are part of NLM’s efforts to make prescription pill matching software freely accessible to health care providers and the general public. Potential public safety benefits include the reduction of medication errors and improvements in poison control and in emergency response.

NLM seeks to better understand incidents involving unidentified or misidentified pills and their impact. Health professionals, educators, law enforcement, private citizens, other affected parties, and RFI respondents are invited to send descriptions of their professional or personal experiences, or links to specifically relevant media coverage, to PIR@nlm.nih.gov.

THE RxIMAGE DATABASE

The RxIMAGE database underlies this RFI and, later, the Challenge and potentially any future pill image recognition apps. This database contains freely available, high quality images and associated data for prescription pills marketed in the United States. It is a growing database that now has images of more than 3,000 prescription pills. Examples of pills are capsules and tablets intended for oral use.

Photos of pills for the RxIMAGE database were taken under laboratory lighting conditions, from a camera directly above the front and the back faces of the pill, and at high resolution. The JPEG images in this database were captured using specialized digital macro-photography techniques. Image files in the RxIMAGE database contain NLM watermarks.

The data associated with each pill image include appearance (color, shape, size, text on the pill, etc.), identifiers such as its National Drug Code[1] (NDC), and ingredients.

THE RFI’s REFERENCE IMAGES

The RFI’s reference images consist of 1,000 JPEG images taken from the RxIMAGE database. While an image in the RxIMAGE database shows both the front and the back of a pill, a reference image is of either the front or the back.

THE RFI’s CONSUMER IMAGES

The RFI’s consumer images consist of 3,000 JPEG images of the same pills that were photographed to create the RFI reference images. However, they were taken with a variety of digital cameras, under various lighting conditions, and at camera angles not necessarily perpendicular to the faces of the pills. They are like photos of unknown prescription pills that the general public might take.

Each RFI consumer image matches one RFI reference image. Conversely, each RFI reference image matches at least two RFI consumer images, taken under different conditions.

SOFTWARE TO BE DEVELOPED FOR THIS RFI

RFI respondents are asked to develop algorithms and software for matching images of consumer pills to reference images – and to submit their executable software as described below. The software does not need to identify pills by name. The software can use any programming language(s).

The software should be a batch-mode program or a script whose input consists of a set of unknown images and a set of reference images. The output should be a comma-separated-value similarity matrix S in which the term s(i,j) is a number in the unit interval [0,1] that indicates how well the i-th unknown image matches the j-th reference image. A larger value should indicate a better match.

Respondents can use the RxIMAGE database and the RFI reference and consumer images in developing their algorithms and software and in testing how well their software works. They can also supplement the consumer images.

Technical details about the reference and consumer images to be made available to respondents, instructions about the software and output that respondents are asked to submit, and the address for submitting a USB drive containing this software are at //pir.nlm.nih.gov

HOW NLM WILL USE THE RFI SUBMISSIONS

NLM will use these submissions to test, evaluate, and as needed refine the components of the forthcoming Challenge: instructions, hardware and software systems, and metrics for use in selecting the winner. RFI submissions will thereby help ensure that the Challenge instructions are accurate and clear, that the Challenge judging will be impartial, and that Challenge responses will include high quality software. The software and output submitted in response to this RFI will not themselves be judged.

The metrics used to judge the Challenge will give high marks to software that correctly matches the most pills. One metric may be the number of images that the similarity matrix says are most similar to their reference images. The metrics may also address what might be called false positives (the similarity is highest but not to the correct reference image) and false negatives (the similarity to the correct reference image is low), the need to consider images of both the front and the back of a pill, color similarity, an unknown image that is about equally similar to multiple reference images, the value of selecting several candidate matches, and speed.

NLM will post information about the responses to this RFI at //pir.nlm.nih.gov. This information may include summary statistics and a discussion of the responses.

DIFFERENCES BETWEEN THIS RFI AND THE CHALLENGE

The Challenge will have its own reference images and its own consumer images. To compete in the Challenge, a person or team must submit a USB drive that contains the algorithm in pseudo-code, the open source human- and machine-readable software that implements the algorithm, the executable code, and the similarity matrix.

Persons and organizations that do not include algorithms and/or open-source code are nonetheless encouraged to participate in the Challenge, but are not eligible to win the Challenge prize. U.S. Government agencies and employees must follow their agency guidelines on Challenge participation.

NLM plans to use the RFI and the Challenge submissions in developing, independently or possibly with others, freely accessible pill image recognition software and tools, such as mobile apps.

RFI SCHEDULE

Date Item
February 2, 2015 Consumer image files and reference image files made available*
April 6, 2015 Start date for accepting responses to this RFI*
May 15, 2015 End date for accepting responses to this RFI

Information about the responses will be made available no earlier than June 22, 2015.

QUESTIONS AND COMMENTS

Questions and comments about this RFI are welcome at PIR@nlm.nih.gov.

MORE ABOUT C3PI

C3PI is a research and development (R&D) project in the Office of High Performance Computing and Communications (OHPCC) within NLM’s Lister Hill National Center for Biomedical Communications (LHNCBC). C3PI computer scientists conduct computer vision R&D in text- and image-based search and retrieval. C3PI’s overall goal is to improve the prescription drug information made available to health care providers and consumers.

DISCLAIMER AND IMPORTANT NOTES

This notice does not obligate the Government to award a contract or otherwise pay for the information provided in response. The Government reserves the right to use information provided by respondents for any purpose deemed necessary and legally appropriate. Any organization responding to this notice should ensure that its response is complete and sufficiently detailed. Information provided will be used to assess tradeoffs and alternatives available for the potential requirement and may lead to the development of a solicitation. Respondents are advised that the Government is under no obligation to acknowledge receipt of the information received or provide feedback to respondents with respect to any information submitted.

CONFIDENTIALITY

No proprietary, classified, confidential, or sensitive information should be included in your response. The Government reserves the right to use any non-proprietary technical information in any resultant solicitation(s).

The National Library of Medicine (NLM) (//www.nlm.nih.gov) is the world's largest library of the health sciences and a component of the National Institutes of Health (NIH) (//www.nih.gov). NLM collects, organizes, and makes available biomedical science information to scientists, health professionals, and the public.

###

[1] //www.fda.gov/Drugs/InformationOnDrugs/ucm142438.htm