Search

• The prediction bias is mainly caused by data imbalance. • Curriculum coefficients are useful for mining hard samples and improving model generalization. • Teacher-student network ensures unbiased glaucoma diagnoses and evidence map predictions. • The evidence map output by the student network can provide more detailed information. Glaucoma is a disease that causes irreversible damage to the optic nerve. Research on accurate automatic screening algorithms is essential for the prevention and treatment of glaucoma. However, due to the imbalance of existing datasets and the existence of some hard samples that accompany other diverse and complex fundus diseases, the performance of current glaucoma screening algorithms is limited. In addition, the lack of interpretability also makes it difficult for the current algorithms to meet the requirements of clinical applications. In this paper, we propose a new multitask curriculum learning framework (MTCLF) for unbiased glaucoma screenings and visualizations of model decision-making areas. MTCLF is a teacher-student framework. The teacher network is used to generate the label evidence map. The student network can diagnose glaucoma and predict the evidence map at the same time with the well-designed dual-branch CNN structure and collaborative learning module. We design two curriculum coefficients θ and σ to guide the training process of the student network in the sample space so that the student network can adaptively balance the sample contribution, reduce the prediction bias and mine hard samples. The experimental results show that the accuracy, sensitivity, specificity, AUC and F 2 -score of MTCLF based on the LAG dataset for glaucoma diagnoses are 0.967, 0.961, 0.970, 0.996, and 0.958, respectively. These results are superior to those of the state-of-the-art methods. MTCLF not only achieves the best performance for unbiased glaucoma diagnoses but also generates a reliable evidence map to help clinicians explore fine lesion areas. [ABSTRACT FROM AUTHOR]

• Dilated convolutions ensure better feature transfer and accurate classification that result in a sensitivity boost. • Dilated convolutions serve as the main constituents of decoder, encoder blocks and skip connections. • Dilated convolutions of different rates are more effective at capturing dynamically complex blood vessels. • Proposed Weighted multi-output fusion retrieves high fidelity vascular map by extracting critical features from each output block. Retinal image segmentation can help clinicians detect pathological disorders by studying changes in retinal blood vessels. This early detection can help prevent blindness and many other vision impairments. So far, several supervised and unsupervised methods have been proposed for the task of automatic blood vessel segmentation. However, the sensitivity and the robustness of these methods can be improved by correctly classifying more vessel pixels. We proposed an automatic, retinal blood vessel segmentation method based on the U-net architecture. This end-to-end framework utilizes preprocessing and a data augmentation pipeline for training. The architecture utilizes multiscale input and multioutput modules with improved skip connections and the correct use of dilated convolutions for effective feature extraction. In multiscale input, the input image is scaled down and concatenated with the output of convolutional blocks at different points in the encoder path to ensure the feature transfer of the original image. The multioutput module obtains upsampled outputs from each decoder block that are combined to obtain the final output. Skip paths connect each encoder block with the corresponding decoder block, and the whole architecture utilizes different dilation rates to improve the overall feature extraction. The proposed method achieved an accuracy: of 0.9680, 0.9694, and 0.9701; sensitivity of 0.8837, 0.8263, and 0.8713; and Intersection Over Union (IOU) of 0.8698, 0.7951, and 0.8184 with three publicly available datasets: DRIVE, STARE, and CHASE, respectively. An ablation study is performed to show the contribution of each proposed module and technique. The evaluation metrics revealed that the performance of the proposed method is higher than that of the original U-net and other U-net-based architectures, as well as many other state-of-the-art segmentation techniques, and that the proposed method is robust to noise. [ABSTRACT FROM AUTHOR]