STARE Diagnostic Approach
The system takes as input an annotation of 39 manifestations. Each manifestation has between two and seven states. For example, the manifestation RPED is considered in the two states absent or present, while the manifestation artery color is considered in the three states normal, copper, or silver. The complete list of manifestations and states is tabulated here for reference.
The output of the system is a set of one to three diagnoses, from a set of thirteen possible diagnoses. For example, the output may be Coats' disease, or Coats' disease AND arteriosclerotic retinopathy.
Our database contains 402 images. A list of ground truth diagnoses for these images is tabulated here for reference. In this file the numbers 1 to 13 are used to denote the diagnoses known to the system, the number 14 is used to denote any diagnosis outside this set of 13 diagnoses, and the number 0 is used to denote the normal diagnosis. Based on these distinctions, lists of familiar, unfamiliar and normal images are tabulated here.
The system utilizes a set of beliefs given by an ophthalmologist. The beliefs are probabilities of the presence of a manifestation in a particular state, given that a disease is present. For instance, the probability that artery color is in the state copper, given that a person has arteriosclerotic retinopathy (ASR), was given as 0.4 (40%). A single entry in the belief table is shown below as an example.
| Manifestation | State | D1 | D2 | D3 | D4 | D5 | D6 | D7 | D8 | D9 | D10 | D11 | D12 | D13 |
| RPED | Absent | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 0.666667 |
| Present | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0.333333 |
The complete set of given belief probabilities is tabulated here for reference. It is also available in Excel spreadsheet format here.
The causal relationships between diseases and manifestations may be represented using a graph, where nodes represent diseases and manifestations, and links represent the causal relationships. The complete set of causal relationships for our system are tabulated here for reference.
A hand drafted annotation of the manifestations in each image was obtained from the expert. The annotations for our 402 images are available here. When the annotations were obtained, the sets of diagnoses and manifestations under consideration were slightly different from the final sets currently in use. Each annotation contains 44 manifestations, compared to the 39 manifestations currently used. A figure detailing these two formats and their differences is available here for reference. The differences may be summarized as follows. The original annotation includes states for retinitis (16), optic nerve in picture (17), retinal angioma (38), chorioretinal scar (40), grape clusters (42), nevus (43) and geographic angioma (44), which are not currently used. The manifestation ON color (21) was split into ON palor (17) and ON color (18). This lets the states for this abnormality be listed monotonically increasing in severity, which is a requirement for the noisy max formulation. Similarly, the manifestation artery color (23) was split into artery oxygen (20) and artery color (21), the manifestation artery diameter (26) was split into artery narrow (24) and artery dilation (25) and the manifestation vein diameter (27) was split into vein narrow (26) and vein dilation (27). The manifestations small or medium blot hemorrhage (11) and retinal hemorrhage (12) were combined to form the manifestation small or medium blot hemorrhage (11). Similarly the manifestations retinal or subretinal exudate (13) and circinate pattern (14) were combined to form the manifestation retinal or subretinal exudate (12).
It would seem natural to apply Bayes' Rule to the problem. However, strict application of Bayes' Rule has several drawbacks: