In medRxiv : the preprint server for health sciences
OBJECTIVE : Poor risk stratification of thyroid nodules by ultrasound has motivated the need for a deep learning-based approach for nodule segmentation. This paper demonstrates the effectiveness of a multitask approach to detect ultrasounds containing potential nodules and segment nodules on the suspected images.
METHODS : Ultrasound studies were acquired from 280 patients at UCLA Health, totaling 9888 images. A novel anomaly detection (AD) module, to classify suspicious ultrasound images, was integrated with various state of the art segmentation architectures. The trained models were evaluated on a portion of the in-house dataset, as well as two external validation (EV) sets, to understand how the AD module affected segmentation performance.
RESULTS : The addition of AD to the architectures improved image-level nodule detection, evidenced by the increase in F1 scores and image-wide Dice similarity coefficient. Of the models with AD, MSUNet-AD had the highest F1 score of 0.829; however, there was a decrease in DSC on just images with nodules ( DSC + ) from 0.726 to 0.627. This drop was observed across all models when AD was added; however, closer analysis of DSC + by nodule size revealed that this difference was not significant in larger nodules, which are more likely to be clinically relevant. Finally, evaluation of MSUNet with and without AD on the EV datasets demonstrated comparable performance with the UCLA dataset.
CONCLUSION : The proposed architecture is an automated multitask method that can both detect and segment nodules in ultrasound. Performance on the EV datasets demonstrates generalizability of the model.
Radhachandran Ashwath, Kinzel Adam, Chen Joseph, Sant Vivek, Patel Maitraya, Masamed Rinat, Arnold Corey W, Speier William
2023-Feb-02
