Subconjunctival Hemorrhage Assisted Diagnostic Based on Attribute Recognition

Subconjunctival hemorrhage is a common ophthalmic disease. It usually does not cause serious visual impairment, but it may be an early sign of serious diseases such as leukemia, hemophilia and septicemia. Traditional manual diagnosis relies on the subjective judgment of doctors. The traditional digital image processing algorithm is only used to classify multiple eye diseases and is not targeted. This paper proposes an auxiliary diagnostic algorithm for subconjunctival hemorrhage based on attribute recognition. The algorithm processes and analyzes a dataset that has been annotated under the guidance of ophthalmologists. By introducing an attention mechanism to the original ResNet50, it enhances focus on effective spatial locations and semantic information channels. Using a multi-task joint training approach, the network can simultaneously recognize nine attributes of subconjunctival hemorrhage, obtaining specific symptoms. To enhance the overall recognition accuracy of the multi-task network, a weighted penalty mechanism is introduced into the loss function. Validation on the test set shows that our algorithm performs exceptionally well in recognizing the nine attributes, with an average accuracy of 92.87% and a highest accuracy of 98.96%.