We contribute in saving the lives of cancer patients through early detection and diagnosis, since one of the major challenges in cancer treatment is that patients are diagnosed at very late stages when appropriate medical interventions become less effective and full curative treatment is no longer achievable. Cancer classification using gene expressions is extremely challenging given the complexity and high dimensionality of the data. Current classification methods typically rely on samples collected from a single tissue type and perform a prerequisite of gene feature selection to avoid processing the full set of genes. These methods fall short in taking advantage of genome-wide next generation sequencing technologies which provide a snapshot of the whole transcriptome rather than a predetermined subset of genes. We propose a Deep Learning framework for cancer diagnosis by developing a multi-tissue cancer classifier based on whole-transcriptome gene expressions collected from multiple tumor types covering multiple organ sites. We introduce a new Convolutional Neural Network architecture called Gene eXpression Network (GeneXNet), which is specifically designed to address the complex nature of gene expressions. Our proposed GeneXNet provides capabilities of detecting genetic alterations driving cancer progression by learning genomic signatures across multiple tissue types without requiring the prerequisite of gene feature selection. We design an end-to-end Deep Reinforcement Learning framework that automatically learns the optimal network architecture together with the associated optimal hyperparameters that maximizes the performance of our multi-tissue cancer classifier. Our framework eliminates the manual process of handcrafting the design of deep network architectures and the manual process of hyperparameter optimization on the target dataset. Our model achieves 98.9% classification accuracy on human samples representing 33 different cancer tumor types across 26 organ sites. We demonstrate how our model can be used for transfer learning to build classifiers for tumors lacking sufficient samples to be trained independently. We contribute in providing medical professionals with more confidence in using Deep Learning for medical diagnosis by introducing visualization procedures to provide biological insight on how our network is performing classification across multiple tumors. To our knowledge, this is the first effort to develop a multi-tissue cancer classifier based on a full set of whole-transcriptome gene expressions collected from tumors across different tissue types without requiring a prerequisite process of gene feature selection.


Computer Science & Engineering Department

Degree Name

PhD in Applied Science

Graduation Date

Winter 1-31-2021

Submission Date


First Advisor

Mohamed Moustafa

Second Advisor

Ahmed Rafea

Committee Member 1

Amir Atiya

Committee Member 2

Hazem Abbas

Committee Member 3

Amr Goneid, Ahmed Moustafa


136 p.

Document Type

Doctoral Dissertation

Institutional Review Board (IRB) Approval

Not necessary for this item