Need help understanding diagnostic criteria for diabetic retinopathy

[throwaway222]

Full disclosure: I am not a medical trainee, but I am considering entering a Kaggle competition designed to help improve automated diabetic retinopathy detection. They are using the International Clinical Disease Severity Scale for Diabetic Retinopathy, and I am trying to understand the criteria for this scale by reading this paper. (Pointers to better tutorials are very welcome.) The paper says

The diagnosis of severe NPDR (nonproliferative DR) is based on the 4:2:1 rule of the ETDRS[20]. Using standard photographs 2A, 6A and 8A to compare with the fundus findings, one can easily diagnose severe NPDR[20]. If hemorrhages of at least the magnitude of standard photograph 2A are present in all 4 quadrants, then by definition severe NPDR is present. If 2 quadrants or more have venous beading (VB) of the same magnitude or greater than standard photograph 6A, then by definition severe NPDR is present. If one or more quadrant has intraretinal microvascular abnormalities (IRMA) of the same magnitude or greater than standard photograph 8A, then by definition severe NPDR is present.

The "standard photographs" are given in Figure 1,

Standard photograph. A: 2A. Notice the intraretinal hemorrhages. If 4 quadrants have intraretinal hemorrhages of at least this magnitude then by definition severe non-proliferative retinopathy is present; B: 6A. Notice venous beading (VB). If 2 quadrants or more have VB of at least this magnitude then by definition severe non-proliferative retinopathy is present; C: 8A. Notice the intraretinal microvascular abnormalities (IRMA). If one or more quadrants has IRMA of at least this magnitude then by definition severe non-proliferative retinopathy is present.

but they are pretty unclear to me. I think I see what's going on in subfigures A & B, but C is a complete mystery to me. For all three photos, versions with more precise pointers to the relevant visual features would be a huge help to me. I.e. this is one of the hemorrhages in subfigure A, distinguished by these characteristics, this is a venous bead in subfigure B, this is an IRMA in subfigure C. Since they are standard photos, I assume someone has done this?

I have found other photos depicting each of these features elsewhere, e.g., these slides have some clear examples of each. But I'm having a real problem mapping those back to the "standard photos" with any confidence, and would be grateful for any assistance in doing so.

---

Members don't see this ad.

 

[MstaKing10]

These are standard photographs from the pivotal government sponsored trial regarding diabetes and its treatment (ETDRS). As such, they have served as the standard by which diabetes is graded for decades in both trials and clinical practice. The first slide shows intratetinal heme and MAs. The second is a great example of venous beading. I've never really liked the last slide representing IRMA but it is the standard. Of course there might be some more dramatic/obvious examples of the above but these are the standard. I would show the pictures to an ophthalmologist and have them show you exactly what all this means, this is difficult to accomplish via posting on a forum

 

[ophthope]

1. Ultimately it's probably going to be hard to learn the information you need and to get input via this forum. I would use this place as a platform to meet Ophthalmologists and get one of them on board with your team. If you're honestly going to code an algorithm to rate fundus photos for diabetic retinopathy you're going to need frequent input and feedback and advice. It seems like a good project and competition.

2. Unless you already have an idea of what the fundus is supposed to look like it will be hard to spot abnormalities. There are a number of picture databases online that may help with that. Honestly if you quickly ran through http://www.ophthobook.com/chapters (specifically ch. 2 and 4) it may give you a good basis. You may have already done some research like this on your own.

3. Intraretinal hemorrhages are the blurry red spots, MAs are tiny anuerysms on the arteries (little more discrete red dots), Venous Beading is when the veins look 'beady' or 'kinked', and IRMAs are hard to describe but are a little more linear instead of dot-like and are confined within the retina as compared to neovascularization which can grow into the vitreous (better picture of an IRMA here). Also just so you know, the link to the slideshow you posted has other types of pictures in it - fluorescein angiography (the black and white photos) that make the pathology more visible. They are different than fundus photos and taking them involves injecting dye into a patient's arm and then taking pictures using certain filters and lights at specific time intervals. They aren't just desaturated fundus photos.


To the other Ophthalmologists here - Can anyone seriously spot an IRMA on these crappy photos? I don't even see an IRMA in photo C. I see DBH and what looks like a cotton wool spot. I can barely tell the difference in real life unless I'm looking at an FA anyway. Does anyone see a tiny red spot in the periphery in the absence of a few quadrants of MAs/heme and say "yep that one spot is more of a little line and it's close enough to an artery that I'm calling it an IRMA and based on the 4-2-1 rule it's time for PRP"? In reality I'm probably calling these things "NVE" anyway. Am I the only one that thinks the distinction is really hard to make without an FA?

 

[MstaKing10]

Irma is really hard to distinguish from nve sometimes. NVE is preretinal as opposed to intratetinal thus above the retina and blood vessels. There may be tufts into the vitreous making it more obvious. Flat nve is really tough to distinguish. Angiogram is the best way in these borderline/tough call cases to know for sure. Recall that Irma does not require immediate PRP, rather, denotes severe npdr and high risk for progression.

I've always disliked the Irma standard photo shown.

 

[throwaway222]

Mstaking10, ophthope, this is very helpful. Thanks! I'll let you know how we do.

 

[Dusn]

I think the goal should be to identify diabetic retinopathy vs no diabetic retinopathy. That'll be easier and more practical in terms of management. Then all the computer has to do is differentiate a red spot of blood caused by a hemorrhage or microanuerysm from a red spot from a normal blood vessel. This would increase the sensitivity and decrease the specificity of the automated algorithm-- which is how a screening test should be. If they have any diabetic retinopathy they need to see an ophthalmologist in order to grade it and talk to the patient about blood sugar control and treatment.

I'd be a little surprised there aren't programs that can do a rudimentary job of this already, as I've seen projects working up more complex analysis of retinal images. Duke is one of the major centers for retina image analysis working on things like this.

 

[DrZeke]

I think the goal should be to identify diabetic retinopathy vs no diabetic retinopathy. That'll be easier and more practical in terms of management. Then all the computer has to do is differentiate a red spot of blood caused by a hemorrhage or microanuerysm from a red spot from a normal blood vessel. This would increase the sensitivity and decrease the specificity of the automated algorithm-- which is how a screening test should be. If they have any diabetic retinopathy they need to see an ophthalmologist in order to grade it and talk to the patient about blood sugar control and treatment.

I'd be a little surprised there aren't programs that can do a rudimentary job of this already, as I've seen projects working up more complex analysis of retinal images. Duke is one of the major centers for retina image analysis working on things like this.

Agreed. None vs any is much more feasible considering the quality of some images.