Direct Prediction of 48 Month Survival Status in Patients with Uveal Melanoma Using Deep Learning and Digital Cytopathology Images.

backgroundUveal melanoma (UM) is the most common primary intraocular malignancy in adults. The median overall survival time for patients who develop metastasis is approximately one year. In this study, we aim to leverage deep learning (DL) techniques to analyze digital cytopathology images and directly predict the 48 month survival status on a patient level.

methodsFine-needle aspiration biopsy (FNAB) of the tumor was performed in each patient diagnosed with UM. The cell aspirate was smeared on a glass slide and stained with H&E. Each slide then underwent whole-slide scanning. Within each whole-slide image, regions of interest (ROIs) with UM cells were automatically extracted. Each ROI was converted into super pixels, and the super pixels were automatically detected, segmented and annotated as "tumor cell" or "background" using DL. Cell-level features were extracted from the segmented tumor cells. The cell-level features were aggregated into slide-level features which were learned by a fully connected layer in an artificial neural network, and the patient survival status was predicted directly from the slide-level features. The data were partitioned at the patient level (78% training and 22% testing). Our DL model was trained to perform the binary prediction of yes-versus-no survival by Month 48. The ground truth for patient survival was established via a retrospective chart review.

resultsA total of 74 patients were included in this study (43% female; mean age at the time of diagnosis: 61.8 ± 11.6 years), and 207,260 unique ROIs were generated for model training and testing. By Month 48 after diagnosis, 18 patients (24%) died from UM metastasis. Our hold-out test set contained 16 patients, where 6 patients had passed away and 10 patients were alive at Month 48. When using a sensitivity threshold of 80% in predicting UM-specific death by Month 48, our model achieved an overall accuracy of 75%. Within the subgroup of patients who died by Month 48, our model achieved a prediction accuracy of 83%. Of note, one patient in our test set was a clinical surprise, namely death by Month 48 despite having a GEP class 1A tumor, which typically portends a good prognosis. Our model correctly predicted this clinical surprise as well.

conclusionsOur DL model was able to predict the Month 48 survival status directly from digital cytopathology images obtained from FNABs of UM tumors with reasonably robust performance. This approach, if validated prospectively, could serve as an alternative survival prediction tool for patients with UM to whom GEP is not available.