Gene expression profile prediction in uveal melanoma using deep learning

Author: T. Y. Alvin Liu (United States)

Co-authors: Haomin Chen, Mathias Unberath, Zelia Correa

Purpose

Uveal melanoma (UM) is the most common intraocular tumor in adults, and the current gold standard prognostic test for metastasis and survival prediction is the gene expression profile (GEP) test [Castle Biosciences, Friendswood, TX]. UM patients can be divided into two classes by GEP, and there is a stark contrast in long-term survival between the two classes: the 92-month survival probability in class 1 patients is 95%, versus 31% in class 2 patients. However, the GEP test is costly and not available outside of the U.S. We aimed to develop a deep learning based algorithm to directly predict GEP status from H&E stained cytopathology slides.

Setting/Venue

Consecutive patients with uveal melanoma treated at a tertiary academic referral center.

Methods

Digital cytopathology slide images were generated from fine needle aspiration biopsies (FNAB) of UM tumors, smearing of tumor cells on glass slides, H&E staining and whole slide scanning at 40x magnification. Our training dataset included 65 slides from 58 patients (28 class 1 tumors; 30 class 2 tumors). Using a customized tool that combined unsupervised clustering and further refinements through human-computer interactions, 90,316 unique regions of interest (ROIs) containing UM cells were generated and used for training of a deep learning system (DLS). GEP status of each tumor was used as ground truth. The data label at the slide level was propagated to the its respective ROIs. The DLS was trained in a two-stage manner, involving first the ResNet-50 deep convolutional neural network (DCNN) and finally a two-layer artificial neural network for the binary classification task of distinguishing between GEP class 1 and class 2. The DLS was tested on a separate dataset of 24 slides from 24 patients (12 class 1 tumors; 12 class 2 tumors; 1 slide per patient).

Results

Our DLS correctly predicted the GEP status of 22/24 slides, thus achieving a 91.7% accuracy on a patient-level analysis. Of the two slides that were incorrectly predicted, one was generated from a class 1 tumor and one was generated from a class 2 tumor.

Conlusions

We have developed a DLS that is capable of robust predictions of GEP status based on H&E stained cytopathology slides alone. Such an algorithm has the potential to serve as an alternative to GEP testing, especially in areas of the world where GEP testing is not available. As the next step, we plan to prospectively validate this algorithm using data from a different institution and develop a similar DLS using actual metastasis status as the ground truth for training and testing.

Financial Disclosure

Amaros (consultant)

Comments

-