A Multitask Learning Framework for Automated Cobb Angle Estimation

Automated Cobb angle estimation on X-ray images is important for scoliosis diagnosis, treatment, and progression surveillance. The noise in X-ray images and the problem of unbalanced samples are the main difficulties of automated Cobb angle estimation, and it is challenging to alleviate error accumulation based on the two-stage cascaded framework. To address these problems, we propose a multitask framework named MSW-Net with two main components—a feature selection module and a weighted discriminant loss—for automated Cobb angle estimation in this article. To reduce the impact of noise in X-rays, we replace the previous spine landmark regression task with the spine segmentation task to provide spine morphology information for the estimation of the Cobb angle. Furthermore, to address the error accumulation problem in the previous two-stage cascaded framework, we propose a joint deep-learning framework for the spine segmentation task and the Cobb angle regression task. Meanwhile, a feature selection module is integrated into this framework to effectively transfer beneficial features between tasks, boosting the overall performance of the joint deep-learning framework. In addition, we also propose a weighted discriminant loss to mitigate the impact of the unbalanced sample, assigning different weights to the loss function according to the number of different deformities. The experimental results on the AASCE dataset show that our proposed multitask framework (MSW-Net) achieves more precise Cobb angle estimation performance than previous methods. Therefore, our proposed MSW-Net can provide reliable, effective, and automated Cobb angle estimation assistance to clinicians.