Pathological diagnosis of thyroid nodules directly from ultrasonography by a weakly supervised deep learning framework.

Background: The management of ultrasonography (US) of thyroid nodules is often time consuming and may be inconsistent between observers. The use of deep learning (DL) can increase the speed of examinations and improve reporting consistency. To develop a novel dual attention-guided DL framework to inference histological status directly from thyroid US without expensive image-level annotations. Methods: In this retrospective analysis, patients first underwent US followed by histopathological examination by fine needle aspiration or surgical histological analysis. Two datasets from the Second Affiliated Hospital of Jiaxing University, Batch 1 and Batch 2, were used in our study for training, testing and validation of the algorithm, respectively. The proposed DL framework is named ThyUS2Path and consists of two attention modules which focus on discriminative nodule patterns in different dimensions. We compare ThyUS2Path with two commonly used state-of-the-art MIL-based methods. Results: We used 6014 images from 603 patients to build the DL framework and evaluated it on 1978 external images, five-fold cross validation AUCs of 0.754 ± 0.035 and 0.735 ± 0.029, respectively. ThyUS2Path outperformed two commonly state-of-the-art MIL-based methods significantly, namely Maxpool and Meanpool. The algorithms also obtained good prediction performance in the external test set (AUROCs of 0.70~0.80 and AUPRCs of 0.78~0.83). Conclusion: Our approach gives a feasible way to link the US phenotype with histological reports and has the potential to augment clinicians' capabilities in thyroid cancer diagnosis in a non-invasive manner.