I'm searching for a topic of interest in the domain of machine learning and computer vision. More specifically, researching can computer vision be applied to classify medical image scans and/or predict the future state of a scan. I'm not a health care professional so in order to frame the problem into something that is attainable I'm aiming to research the topic of image scanning some more.

What I would like to know is :

What is the current state of the art of image scanning technology ?

What are it's weaknesses ?

I have many unknown unknowns and I'm unsure where to start to attain a basic knowledge.

Book recommendations are welcome, for example this book seems like a good place to start : For example the book "Medical Imaging for the Health Care Provider: Practical Radiograph Interpretation " : https://www.amazon.com/Medical-Imaging-Health-Care-Provider-ebook/dp/B01HUNOJPG

The dataset I'm intending to use for this research is 'DeepLesion' https://www.nih.gov/news-events/news-releases/nih-clinical-center-releases-dataset-32000-ct-images

Update : this seems like a good place to start : Medical imaging - image quality?

Update 2:

I'm aiming to utilize image and annotated data from DeepLesion to develop an AI to diagnose future and/or present state of a scan. 'future state of a scan' refers to predicting the future state of scan attributes . The attributes are what is contained in the DeepLesion annoted dataset which includes lesion diameter, patient gender and patient age. So I will attempt to predict 1 or a combination of these attributes. 

At this stage I'm not aiming for the AI model to perform a diagnosisis or prognosis but provide a prediction of attributes that aids the healthcare practitioner in performing the diagnosis or prognosis. Due to DeepLesion containing CT image scans the healthcare practitioner in this case is a radiographer. 

Other type of more high level predictions/classifications I may consider are detecting liver, lung, kidney lesions.

The type of predictions are dependant on the type of data available. 

Another research question I have is what type of predictions are of most value to the practitioner. This will help focus my research.

  • 2
    A little confusing to understand what you're asking... is it having an AI read an image and return a diagnosis and/or prognosis? The trajectory of that technology is unclear as of yet but promising. But if so, I'm concerned this question could be exceedingly broad in scope to be able to answer, Also, it's far less likely to get much of a response on this site, as this is more about clinical medicine than biotechnology...
    – DoctorWhom
    Jan 5, 2019 at 5:55
  • 1
    What is your exact goal? Are you interested in developing computer hardware or software that would help in image diagnostics? Can you explain this bit: "can computer vision be applied to classify medical image scans and/or predict the future state of a scan. "? What do you mean by "the future state of a scan?"
    – Jan
    Jan 5, 2019 at 9:19
  • Have you also tried the AI SE site? Jan 5, 2019 at 10:20
  • @DoctorWhom please see question update 2.
    – blue-sky
    Jan 5, 2019 at 15:54
  • 1
    Can you make a practical example of predicting the future state of the scan? I imagine this involves, for example, looking at a certain CT image and then use computer vision that can tell more than what you can see with your own eyes? But it would be much more clear if you make an exact example with exact attributes.
    – Jan
    Jan 5, 2019 at 16:17

2 Answers 2


Commonly used diagnostic imaging methods are:

  • Ultrasonography
  • X-ray
  • Computed tomography (CT)
  • Magnetic resonance imaging (MRI)
  • Scintigraphy or radionuclide scan (injecting a radioactive tracer into a vein, waiting until it collects in a certain organ, for example, the thyroid gland, and taking a picture of the tracer distribution with a scanner)

All the mentioned techniques have several variants, for example Doppler ultrasound, an MRI with contrast, etc. The Wikipedia's Medical Imaging has a more detailed "index" of techiques with links to individual articles.

On Biology SE, there's a list of websites that provide open-access images, some of which come with the descriptions of cases. Before buying any book, I strongly recommend you to get a clear idea about which types of books can serve your purpose. A book that can be excellent for a doctor or medical student can be useless for you. I also recommend that, for a start, you chose ONE imaging technique and research it a bit, rather than go with all imaging at once; the problems in ultrasonography are significantly different than in CT.

Examples of weaknesses of imaging techniques:

  • A CT and MRI, at least, are expensive.
  • An X-ray can show only lesions that are significantly more or less radiopaque than the surrounding tissues (for example, it can show only gallstones rich in calcium but not others).
  • An MRI of the gallbladder cannot reliably distinguish between the noncancerous polyps and cancers (Radio Graphics).
  • The most common problem is probably, that despite high sensitivity (ability to detect a lesion), specificity (ability to reveal/predict an exact type of the lesion) of CT and MRI scans can still be low.

A common question for a health practitioner that often remains unresolved after imaging is: Is the lesion cancerous or not or how likely will it develop into a cancer. For example, gallbladder polyps greater than 10 mm are significantly more likely cancerous than the smaller ones, but it is not clear if the risk increases after 5 mm or after 15 mm, for example. Also, sometimes imaging fails to show if cancer has spread to nearby organs.


To predict a lesion on a CT image, you need to know how a normal CT image looks and how a lesion looks. The knowledge about what is a lesion came from comparisons of many CT images and the actual physical situations discovered during surgery. Now, to extend this knowledge beyond what you can see with your own eyes on a CT image, you would again need to compare many CT images (using a computer program) with surgery results.

I imagine, this would require a project in which several experienced radiologists, surgeons and computer experts would be involved. One project would need to focus on a single question, for example: What are predictors of gallbladder cancer in abnormal gallbladder growths detected on a CT image? Thousands of CT images and surgery results would then need to be compared to find eventual associations.


I would like to make a recommendation, as a researcher also working in medical imaging. You state that you are interested in predicting lesion diameter, patient gender, and patient age from the scans. However, when a radiologist reads a scan, they already know the patient gender and the patient age because that information is in the medical record. They also know the reason for the scan. For example, they will often see a display like "Ms. Smith is a 55-year old woman with a history of lung cancer" along with the scan itself. (And they can click on the patient's medical record and view everything in the medical record if they want.) So, I think you are better off not predicting things that are already known to the doctor. There are many other cool medical imaging tasks you can do with the DeepLesion data set, e.g. like predicting if there is a lesion in the scan.

Here are some other resources that might help you:

  1. Overview of basic chest anatomy for radiology and abdomen anatomy for radiology
  2. Radiology terms of location
  3. How to read normal chest x-rays. Chest x-rays are not CTs, but if you are just learning about medical imaging it is easier to start with chest x-rays and then move on to CTs
  4. Interpretation of abdominal CTs

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.