Artificial Intelligence in Medical Imaging

We explained in our last article how Artificial Intelligence is doing wonders in the healthcare industry by making diagnosis and treatments easier. We worked on a project in the line of Lung Cancer diagnosis and treatment.

Most lung cancer falls into two categories: Small-cell lung cancer (SCLC) and non-small-cell lung cancer (NSCLC). NSCLC is more prevalent, accounting for approximately 85% of all lung cancers. NSCLC is more complex to treat and the patient’s prognosis can be influenced by obesity and specifically, excess visceral fat. Therefore, it is the characteristics of the host – not the tumor4 – that aids in predicting mortality, treatment planning, and expectation management. The use of the patient’s body mass index (BMI) is insufficient to give an accurate picture of the amount of visceral fat because it does not differentiate between the visceral fat area (VFA) and the subcutaneous fat area (SFA). The SFA is not a risk factor for NSCLC mortality but must be measured separately to calculate the visceral fat Index (VFI). The VFI is expressed as VFI = VFA / SFA + VFA.3 Images from a full-body, low-dose computed tomography (LDCT) scan facilitate the measurement of VFA and SFA2.

Problem Statement

Determining the ratio of visceral fat to subcutaneous fat is vital to physicians in treating patients with non-small cell lung cancer. This is best accomplished by selecting a precise slice from a computed-tomography (CT) scan at the level of the third lumbar vertebra (L3), measuring both types of fat, and calculating the index of visceral fat.

The manual method of identifying the L3 to select and measure the correct slice from the CT image is cumbersome and subjective. Early artificial intelligence (AI) algorithms were accurate but lacked confidence, recommending manual review on ~40% of results. Initially, radiologists used the L4/L5 landmark to select a CT scan slice and measure the VFA from that slice. However, the practice has shifted to pinpointing L3 – rather than L4/L5 – to get a more accurate approximation of the VFA.1

Although the shift to L3 produces more accurate results, this manual method is still subjective. A radiologist can visualize the L3 with more or less clarity, but the margin of error and lack of confidence in the data can be significant.

Why we need an urgent solution?

Better data means better patient outcomes. Transforming VFA measurement from a manual process to an automated process can increase the efficiency and reliability of VFI calculations. Basing the automated process on artificial intelligence (AI) and machine learning (ML) can produce repeatable results and demonstrate a high degree of confidence in the data.

Project Goals

We set out to prove that automating a previously manual process leads to better data and greater confidence in the calculation of the VFI. Specifically, we targeted three project outcomes:

  • Aid radiologists and other physicians in the accurate identification of the L3 vertebra and automated calculation of the VFI
  • Create an algorithm that is successful in identifying L3 slices from full-body CT scans
  • Demonstrate confidence by analyzing results reported by the algorithm as either “yes” (the location of L3 is accurate) or “review” (a manual review is recommended)
  • Overcome the historic lack of confidence in automated calculations by demonstrating a high level of confidence in the calculations made through the use of AI and ML.

We will be discussing our methodology, solution, and achievements through the next weeks. Please follow us on LinkedIn to stay tuned.