22 May 2025 in IEEE Access
Automatic Diabetic Retinopathy classification is crucial as it aids in determining the disease stage, enabling timely treatment to prevent progression to low vision and significantly reducing the diagnostic burden on ophthalmologists. Efficient and accurate network models are essential for clinical applications. In this paper, we propose a deep learning method for DR diagnosis and grading based on multi-scale feature fusion and attention guidance. This approach employs dilated convolutions with varying dilation rates to expand the receptive field and integrate features at multiple scales. Additionally, an attention mechanism based on spatial directional feature aggregation is embedded in the inverted residual modules to guide the model to focus on critical lesion features in fundus images. Furthermore, transformers are utilized to extract global features, enhancing model performance. Experiments conducted on the APTOS-2019 and DDR datasets demonstrate that our method achieves a diagnostic accuracy of 99.2% and grading accuracy of 88.6% on the APTOS-2019 dataset, and a grading accuracy of 84.8% on the DDR dataset. Experimental results show that the proposed method performs excellently and outperforms many recently proposed popular DR diagnosis and classification methods.
9 January 2020 in Journal of Healthcare Engineering
Mosaicking of retinal images is potentially useful for ophthalmologists and computer-aided diagnostic schemes. Vascular bifurcations can be used as features for matching and stitching of retinal images. A fully convolutional network model is employed to segment vascular structures in retinal images to detect vascular bifurcations. Then, bifurcations are extracted as feature points on the vascular mask by a robust and efficient approach. Transformation parameters for stitching can be estimated from the correspondence of vascular bifurcations. The proposed feature detection and mosaic method is evaluated on retinal images of 14 different eyes, 62 retinal images. The proposed method achieves a considerably higher average recall rate of matching for paired images compared with speeded-up robust features and scale-invariant feature transform. The running time of our method was also lower than other methods. Results produced by the proposed method superior to that of AutoStitch, photomerge function in Photoshop cs6 and ICE, demonstrate that accurate matching of detected vascular bifurcations could lead to high-quality mosaic of retinal images.
17 September 2025 in IEEE Access
Diabetic Retinopathy (DR) grading plays a pivotal role in clinical decision-making, where precise staging directly impacts treatment planning and blindness prevention. Inspired by ophthalmologists’ diagnostic reasoning—which systematically integrates local lesion assessment with global pathological pattern analysis while incorporating edge and frequency-domain information for comprehensive image analysis—we propose an innovative hybrid framework. This framework emulates the clinical cognitive process through three synergistic components, constructing a clinically aligned feature hierarchy: (1) A Lightweight Mixed-domain Attention (LMA)-driven biomarker localization module enhances sensitivity to microaneurysms and exudates by learning spatial-channel dependencies while maintaining a lightweight architecture. (2) Frequency Multi-scale Dilated Convolutional Blocks integrate multi-scale spatial and frequency domain features, mimicking the human visual process in lesion identification. (3) Vision Transformers model long-range dependencies to analyze lesion distribution patterns and quantify pathological severity, analogous to clinicians’ holistic evaluation. The entire model contains only 10M parameters, effectively bridging the local modeling capacity of convolutional networks with the global reasoning ability of transformers. Experiments on the APTOS-2019 and DDR datasets demonstrate that our method achieves a grading accuracy of 89.6% and a QuadraticWeighted Kappa (QWK) of 93.1% on APTOS-2019, as well as a grading accuracy of 85.2% and aQWKof 84.9% on the DDR dataset. Results show that the proposed method outperforms many recently popular DR diagnosis and classification approaches, validating its excellence in performance.
29 January 2019 in Computational and Mathematical Methods in Medicine
Automatic segmentation of ulna and radius (UR) in forearm radiographs is a necessary step for single X-ray absorptiometry bone mineral density measurement and diagnosis of osteoporosis. Accurate and robust segmentation of UR is difficult, given the variation in forearms between patients and the nonuniformity intensity in forearm radiographs. In this work, we proposed a practical automatic UR segmentation method through the dynamic programming (DP) algorithm to trace UR contours. Four seed points along four UR diaphysis edges are automatically located in the preprocessed radiographs. Then, the minimum cost paths in a cost map are traced from the seed points through the DP algorithm as UR edges and are merged as the UR contours. The proposed method is quantitatively evaluated using 37 forearm radiographs with manual segmentation results, including 22 normal-exposure and 15 low-exposure radiographs. The average Dice similarity coefficient of our method reached 0.945. The average mean absolute distance between the contours extracted by our method and a radiologist is only 5.04 pixels. The segmentation performance of our method between the normal- and low-exposure radiographs was insignificantly different. Our method was also validated on 105 forearm radiographs acquired under various imaging conditions from several hospitals. The results demonstrated that our method was fairly robust for forearm radiographs of various qualities.