Challenge Description

Building on the opportunity to improve early detection of lung cancer, this community-driven project aims to develop an end-to-end application that connects the predictive power of machine learning with functional software tested against errors and a clean user interface focused on clinical use.

The application will focus on three big challenges that can help radiologists detecting lung cancer in practice:

IDENTIFICATION

Analyze CT scan images to detect and pinpoint the location of concerning nodules from background tissue.

CLASSIFICATION

Use what we’ve seen from nodules in the past to predict whether the identified nodules are cancerous or benign.

SEGMENTATION

Find the boundaries of nodules and create automatic measures to help radiologists refine and build out the computer-aided diagnosis.

The user story

Bob is 55 years old. He goes to his family doctor for a routine checkup. The family doctor notes that Bob fits the right diagnostic profile for getting a lung cancer early screening scan. Bob’s doctor schedules him for a low dose computed tomography (LDCT) scan at a teaching hospital down the street. Bob goes to this appointment a few days later where a technician helps him into the computed tomography (CT) machine and then captures imagery of his chest cavity.

Dr. Smith is a chief resident in diagnostic radiology at the hospital where Bob’s scans were taken, and works in a lab that conducts experimental research using new technologies and treatments. She is an experienced and competent professional with years of experience making diagnoses using standard Picture Archiving and Communication System (PACS) software.

Still, like most radiologists she acknowledges that there is an art to detection and diagnosis and that if several other radiologists examined the same CT scans they might all reach slightly different conclusions.

She also knows that computer aided detection (CADe)/computer aided diagnosis (CADx) technology has been advancing rapidly and she’s open to using an easy-to-use tool that will help her do her job more efficiently.

Dr. Smith has just pulled up Bob’s imagery in order to review and interpret the imagery that the CT scanner generated. In front of Dr. Smith are two computers:

  • To the left is her ordinary hospital workstation running a standard PACS software package. This is a system she uses every day, and she is extremely practiced at using this software; ever since starting residency, she has used some similar version of PACS software nearly every day. The PACS software is a highly specialized and fully featured software suite geared towards making it easy for clinicians to explore imagery. For instance, the user can pan around the image, zoom in and out of specific areas, scroll up and down the Z axis through image layers, change image brightness and contrast, and measure lengths between points.
  • To the right is a standalone laptop. The software from this project is running locally on a fresh Linux install, and a browser window is opened on the homepage of the web service. This computer is trusted and only has intranet access to the DICOM imagery server.

The key question is this: what could the computer on the right enable that Dr. Smith would find valuable enough to use and would make a practical difference for her and her patients?

Here are some issues in the current PACS diagnosis workflow that Dr. Smith feels could be improved upon:

  • Radiologists don’t want to miss any spots with potentially concerning tissue.The current software is great at presenting imagery but doesn’t adequately help detect nodules.
  • Radiologists worry a lot about false positives because exposing patients to additional CT scan radiation or unnecessary surgical procedures carry risks of their own and can end up being more dangerous than the original nodules. The current software doesn’t adequately help minimize false positives.
  • The more accurately a nodule is measured, the more accurate the diagnosis. Right now, most radiologists just eyeball which slice looks like it has the widest diameter. Then they use the measuring tool to take one horizontal measurement in the direction that looks widest. This method is imprecise and results in a loss of useful information when a complex 3D shape having volume and surface features is reduced to a single estimated width. Additionally, not knowing the precise nodule boundaries also means it’s difficult to quantify whether and how a nodule has changed over time. The absence or presence and characteristics of such change over time is a major cancer indicator for radiologists. So the current software doesn’t adequately help measure nodule boundaries.
  • There is an existing set of best practices for what should be included in a lung cancer screening report (i.e. RSNA Standard Template For Lung Cancer Screening, http://www.radreport.org/template/0000268). Most radiologists haven’t adopted this standard yet, partially because it’s a lot of work and partially because the tools aren’t helpful in filling this out. The current software doesn’t adequately help streamline reporting.

Read more about the scope, architecture, and technical requirements of the application in the project's design document.

Next step: making contributions →