Retinal microaneurysm detection based on transformation splicing and multi-context ensemble learning. (April 2022)

• Record Type: Journal Article
• Authors: Du, Jingyu; Zou, Beiji; Ouyang, Pingbo; Zhao, Rongchang
• Is Part Of: Biomedical signal processing and control. Volume 74(2022)
• Publication Date: 2022-04
• Subjects: Microaneurysm Detection -- Data Enhancement -- Local Cross-section Transformation -- Image Splicing -- Multi-context ensemble learning; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2022.103536
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: . A novel data enhancement scheme, transformation splicing (TS), is proposed, overcoming the extreme imbalance of microaneurysm and interference from tiny blood vessels. A multi-context ensemble learning model is designed to adaptively learn multiple image contexts from different spliced images, which improves the detection performance for weak microaneurysms. The proposed microaneurysm detection method achieves state-of-the-art performance on three public datasets. Abstract: Retinal microaneurysm (MA) detection is essential for diagnosis of diabetic retinopathy (DR) by providing the earliest clinical sign of DR. However, automatically detecting MA has always been a challenge due to the extremely small proportion of MA, the susceptibility to interference from blood vessels, and the obvious contrast difference between MAs. This paper proposed a novel deep learning method to achieve accurate MA detection based on transformation splicing (TS) and multi-context ensemble learning. TS rebalances the proportion of MA and reduces interference from blood vessels by transforming the pixel distribution of each candidate image and reinforcing the features of difficult samples, which enables the subsequent model to better learn the enhanced image features. At the same time, a multi-context ensemble learning combining dual deep learning models and attention mechanism is designed to adaptively learn different spliced image contexts, which improves detection performance for weak MAs. The final scores of the proposed method in e-ophatha-MA, DiaretDB1 and ROC three public datasets are 0.518, 0.429 and 0.306 respectively, which demonstrates the state-of-the-art performance for MA detection.