Virtual Medical Centre

Computer Aided Detection (CAD) in Radiology

• Introduction to computer aided detection (CAD) in radiology
• What is mammography and how has CAD helped?
• Imaging by chest CT: Detection of lung nodules and cancer
• Prediction of lung function after an operation

Introduction to computer aided detection (CAD) in radiology

Radiology is a particular field of medicine that uses imaging technology to help doctors diagnose and treat diseases. A radiograph involves exposing a particular part of the body (that is to be imaged) to a small dose of ionising radiation. This produces pictures of the inside of the body.

Over the years, radiological examinations are now increasing in complexity and consist of hundreds of images per study. With this increased complexity, the radiologist's failure to detect an abnormality is higher than previously seen. This is even when an expert radiologist works under ideal conditions. Interpretation errors (errors of diagnosis) are less frequent. This has lead to the use of computers, in what is now known as Computer Aided Detected (CAD) in radiology. A computer can continue to view films and pictures indefinitely, as it does not get tired like humans do.

What is mammography and how has CAD helped?

Mammography involves using low dose X-rays to examine and view the human breast. It is widely used as a screening tool in women for breast cancer. CAD has been shown to be useful in mammography.

In the USA in 1998, an automated software program which could act as a second set of eyes to assist the radiologist to read mammograms was approved by the FDA. The CAD unit sells for US $200,000 but is not 100% sensitive. (Sensitivity is a measure of how well a test identifies the positive cases).

A study performed by radiologists and science officers found that the computer sometimes fails to mark an area the human radiologist thinks may be a cancer. The fact that the CAD did not locate the lesion does not mean there is no cancer and it is very important that the patient is still reviewed and recalled. On the other hand, experienced radiologists have been known to miss cancers.

In a study conducted in 2002, 1000 mammograms were reviewed. These mammograms were from women who had a 'normal looking' earlier mammogram, but who were now diagnosed with breast cancer. 67% of those cancers were visible on the earlier mammogram. Some only had very slight changes, but 27% were visible and required further action (instead of being called normal). This 27% of mammograms were exposed to CAD, and this technology picked up 77% of those cancers, which had previously been described as normal by radiologists.

In the United Kingdom National Breast Screening Program, 10,267 mammograms were obtained in women aged 50 years or older. A single reading with CAD led to 6.5% more cancers being detected than with double reading (ie the mammograms were reviewed twice) by radiologists. However, there was also an increased percentage of women (32%) who were recalled, to come back for further investigation with CAD.

Imaging by chest CT: Detection of lung nodules and cancer

CT scanning (also known as CAT scanning) is a non-invasive type of medical imaging that helps doctors diagnose and treat medical conditions. CT scans use multiple x-rays from special imaging equipment to produce many images of the inside of the body. A computer system then works to join them together.

Chest CT scans allow us to obtain a much more detailed view of the body's internal organs behind the rib cage, such as the lung and also allows views of the ribs, other bony structures and soft tissues.

The first CAD system for CT images of lung was approved in 2004 and 2006 by the FDA. This CAD system was to act as a second reader of chest multi-detector CT scans (MDCT). Their main function was to pick up any abnormalities and alert the radiologist to lesions that they may have been missed in their first reading. Images from a MDCT scanner are received, the CAD unit processes the data and marks regions of interest / abnormality. The radiologist can then perform a second review of the scan, reviewing the areas that the CAD has highlighted.

In 2006, an analysis of 150 low-dose screening CT studies of lung was performed. Nodules (small rounded masses of tissues) seen in the lungs were classified by their size, density and location. CAD detected 72.6% of true nodules and nodules in 4% patients not identified by radiologists. This resulted in a change to the follow-up protocol. A combination of review by the radiologist and by CAD was necessary to identify a total of 1,106 nodules in the 150 patients.

 A lung screening program has found that the rate of overall nodule detection by CAD is 70% sensitive in 10mm thick, low dose, CT scans. When the method was applied and trained specifically to look for malignant nodules, a sensitivity of 80% was attained.

This automated computer detection method needs to also be compared against findings from a radiologist. In addition, software programs that contain information as to whether a nodule is benign or malignant would also be helpful when used in conjunction with CAD.

There was a study using thin CT slices (1mm thick) to compare the characteristics of malignant nodules, compared to benign nodules. 222 cases were studied, consisting of 59 cancers and 163 benign nodules. The results showed that malignant nodules were more often round in shape, and consisted of a mixed texture, made of varying opacities. In the solid nodules, a polygon shape or smoothed margin was seen more often when the lesion was benign.

 However, the true test to determine whether a lesion is benign or malignant is a biopsy. This involves the removal and examination of a sample of tissue directly from the lesion. Another study reviewing high-resolution CT (HRCT) findings of small nodules resected from the lungs was conducted in 2006. 223 nodules which were 2cm or less in diameter were studied. A characteristic appearance was identified in lung cancers known as adenocarcinoma. These types of cancers demonstrated a mixed solid and opaque (non see through) appearance. Indentation of the lining of the lung was found in 75.2% of adenocarcinomas and another type of lung cancer known as squamous cell carcinoma. However, this finding was also found in benign conditions such as inflammation, e.g. tuberculosis, so this feature is not specific for malignancies.

These criteria that the study identified could help distinguish different types of cancers and lesions, and would be helpful if incorporated into the software programs for CAD. This would allow the computer system to provide more information on the likely nature of a lesion.

A five-year study was performed by the Mayo Clinic using CT scans to screen for lung cancer. Their initial results did not support the possibility of reducing the overall deaths from lung cancer. Concern was raised about the false-positive results and overdiagnosis actually resulting in more harm than good. False positive results mean that a lesion might be identified as a cancer or suspicious for a cancerous mass, but in fact, it is not.

CT screening permits detection of a large number of benign lung nodules. These will be expensive to diagnose and may have an impact on general well-being and mortality if we intervene. With computer assisted detection allowing us to identify even more nodules, the benefits and ethical issues must all be considered carefully.

Prediction of lung function after an operation

A software tool developed by Beyer and a team of other researchers allows us to predict the likelihood of operating and expected outcome in some patients with lung cancers. This uses a chest multi-slice CT scanner and results from lung function tests that are performed in the patient prior to the operation.

Digital chest radiography 

A study looked at how useful CAD systems were on 50 cases of lung cancer which were thought to be amenable to operation, compared to 50 normal control patients. Digital chest radiography equipment was used to evaluate the cases. Overall, the CAD system was able to detect 74% of the lung cancer cases.