Hybrid model enabling highly efficient follicular segmentation in thyroid cytopathological whole slide image

Background The prevalence of thyroid cancer is growing rapidly. Early and precise diagnosis is critical in thyroid cancer caring. An automatic thyroid cancer diagnostic tool can be valuable to achieve early detection and diagnostic consistency. Only the follicular areas in the sample contain useful information to the thyroid cancer diagnosis based on fine needle aspiration (FNA). This study aimed to develop a highly efficient accurate method for follicular cell areas segmentation (FCAS) of thyroid cytopathological whole slide images (WSIs).Methods A total of 96 cell samples from July 2017 to July 2018 were collected in one hospital in Beijing, China. Forty-three WSIs were selected and manually labeled, including 17 cases of papillary thyroid carcinoma sample and 26 cases of benign sample. Six thousand and nine hundred cropped typical image patches (available on https://github.com/bupt-ai-cz/Hybrid-Model-Enabling-Highly-Efficient-Follicular-Segmentation) of 1024 × 1024 pixels from 13 large WSIs were used for patch-level model training and testing and all of the 13 large WSIs were papillary thyroid carcinoma samples. Thirty testing WSIs with an average size 36,217 × 29,400 (from 10,240 × 10,240 to 81,920 × 61,440) were used to test the effectiveness of the hybrid model. Based on the traditional semantic segmentation model deeplabv3, we constructed a hybrid segmentation architecture by adding a classification branch into the segmentation scheme to improve efficiency. Accuracy was used to measure the performance of the classification model; pixel accuracy (pAcc), mean accuracy (mAcc), mean intersection over union (mIoU), and frequency weighted intersection over union (fwIoU) were used to measure the performance of the segmentation model, respectively.Results Using this method, up to 93% WSI segmentation time was reduced by skipping the colloidal areas and the blank background areas. The average processing time of 30 WSI was 49.49 s. On the patch dataset, this hybrid model might reach pAcc=98.65%, mAcc=85.60%, mIoU=79.61%, and fwIoU=97.54%. On the WSI dataset, this model might reach pAcc=99.30%, mAcc=68.94%, mIoU=58.21%, and fwIoU=99.50%.Conclusion The proposed hybrid method might significantly improve previous solutions and achieve the superior performance of efficiency and accuracy.