Multi-indices quantification of optic nerve head in fundus image via multitask collaborative learning.

Multi-indices quantification of optic nerve head (ONH), measuring ONH appearance with multiple types of indices simultaneously from fundus images, is the most clinically significant tasks for accurate ONH assessment and ophthalmic disease diagnosis. However, no attempt has been reported due to its challenges of the large variation of fundus appearance across patients, heavy overlap and extremely weak contrast between optic nerve head areas. In this paper, we propose a multitask collaborative learning framework (MCL-Net) for multi-indices ONH quantification. The proposed MCL-Net, a two-branch neural network, first obtains expressive shared and task-specific representations with the backbone network and its two branches; then models the feature exchanges and aggregations between two branches with a well-designed feature interaction module (FIM) to promote each other collaboratively. After that, it estimates multiple types of ONH indices under a multitask ensemble module (MEM) that is capable of learning aggregation of multiple outputs automatically. Therefore, the proposed MCL-Net is consisted of the feature representation, inter-task feature interaction, dual-branch task-specific prediction, and multitask quantification ensemble, which establish an effective framework which takes full advantages of segmentation and estimation tasks for multi-indices ONH quantification. Rather than the low-level feature sharing and individual prediction, the proposed MCL-Net collaboratively learns an optimal combination of shared and task-specific representation, as well as the aggregated prediction, therefore leads to accurate quantification of ONH with multiple types of indices. Experimental results on the dataset of 650 fundus images show that MCL-Net successfully delivers accurate quantification of all the three types of ONH indices, with average mean absolute error of 0.98 ± 0.20, 0.97 ± 0.16, 1.19 ± 0.18, as well as average correlation coefficient of 0.699, 0.708 and 0.691, for diameters, whole areas and regional areas, respectively. In addition, the experiments demonstrate that quantitative indices obtained by our method provide more effective glaucoma diagnosis with AUC of 0.8698. This endows our proposed MCL-Net a great potential in clinical assessment from focal to global for ophthalmic disease diagnosis.