UC Irvine

ABSTRACT : Background: Parotid tumors represent 70-85% of all salivary gland masses and 15% of identified parotid tumors are malignant. Malignant tumors require surgical intervention and more advanced tumors require more extensive surgical resection with a higher likelihood of facial muscle dysfunction, nerve damage, and disfigurement. Automated accurate segmentation of parotid tumors can be a valuable diagnostic aid in busy clinical practices with the potential to facilitate earlier detection and intervention to avoid adverse outcomes. It is also a critical step in advancing computational image analysis, including radiomics and other machine learning workflows. This study proposes a deep learning solution for parotid tumor segmentation and screening. Methods: This study consists of a segmentation task, wherein the algorithm generates an outline of any identified parotid mass, and a screening task, wherein the algorithm assigns a scan to a binary state of containing or not containing a parotid mass. For the segmentation task, a retrospective cohort of patients with parotid masses was aggregated from two separate academic centers. For the screening task, a cohort of consecutive patients was aggregated from a single academic center. All exams were visually inspected for the presence of a parotid mass > 10 mm; when available, histopathology was used to verify diagnosis. 3D tumor masks outlining any masses were generated for each patient by a CAQ-certified neuroradiologist. Both dedicated CT neck protocols and routine exams (including head CT) were included to maximize algorithm generalizability. Two serial 3D deep learning algorithms were developed. The first algorithm localizes the right/left parotid glands individually and is optimized through random sampling of known tumor and non-tumor regions. The second algorithm uses cropped volumes generated by the first algorithm as inputs into a 3D contracting-expanding (U-Net) segmentation model. Both models are implemented using an identical 3D network comprised of 15 convolutional layers and 578,089 parameters. Results: A total of 201 patients with parotid masses were identified from two academic medical centers (N=100 for first site, N=101 from second site). The median tumor volume was 4.62 cm^3 (IQR 2.40-12.50 cm^3). Parotid mass segmentation yielded a Dice score of 0.73 (IQR 0.500-0.788), wherein the Dice score is a metric to measure the pixel similarity between a segmented image and a ground truth that ranges from 0 (no similarity) to 1 (identical). The test performance yielded an AUC of 0.96, accuracy 0.90, sensitivity 0.88, specificity 0.92, PPV 0.93, and NPV 0.87. No significant differences in performance were noted between the different academic centers or imaging protocols (p > 0.05). For the screening task, a total of 200 consecutive unique asymptomatic patients were identified from a single academic center. The binary classification task had a sensitivity of 0.95 and a PPV of 0.90 with a tumor threshold size of 500 pixels. The deep learning model yielded a total 8 positive predictions, 3 of which were confirmed by a neuroradiologist to be true positive parotid masses; none of the masses were identified in the original radiology report. The algorithm yielded a Dice score of 0.65 when evaluating the now total of 401 positive and negative cases. Conclusions: The proposed automated algorithm can accurately: (1) detect incidental parotid masses on routine CT exam; (2) segment parotid tumors for calculation of tumor volume as well as facilitating radiomics and other machine learning workflows. Accordingly, the application of this algorithm in clinical settings has the potential to facilitate earlier detection of parotid tumors.