Session 14 - Healthcare: A Dense-Inception Network for Medical Image Classification

Presentation Type

Oral

Student

Yes

Track

Health Care Application

Abstract

Recently, the advent of deep convolutional neural network (DCNN) have led to impressive accomplishments on image recognition tasks which was preceded by the breakthrough performance of Inception network for solving image and recognition tasks. In healthcare economies, deep learning approaches have shown great promise in medical diagnosis and treatment. In this work, we proposed a new model architecture that integrates DenseBlocks into the Inception Module that we refer to as DINET. The novel model captures multi-scale learned features while increasing parameter efficiency throughout the network. Our aim is to investigate the impact of maintaining a balanced tradeoff between model efficiency and effectiveness of preventing the problem of vanishing gradient. For our medical images classification problem, specifically using Chest x-ray images, experiment results using our DINET model shows modest performance improvements. In our future research directions, other versions of DINET modules will be experimented to shed more light on the impact of strategically positioning the DINET modules in the network to achieve the trade-offs of model efficiency and performance while increasing interaction of multi-scale learned features.

Key words: Chest X-rays, DCNN, Deep learning, DINET, healthcare, medical images.

Start Date

2-11-2020 3:30 PM

End Date

2-11-2020 4:30 PM

This document is currently not available here.

Share

COinS
 
Feb 11th, 3:30 PM Feb 11th, 4:30 PM

Session 14 - Healthcare: A Dense-Inception Network for Medical Image Classification

Campanile & Hobo Day Gallery (A & B)

Recently, the advent of deep convolutional neural network (DCNN) have led to impressive accomplishments on image recognition tasks which was preceded by the breakthrough performance of Inception network for solving image and recognition tasks. In healthcare economies, deep learning approaches have shown great promise in medical diagnosis and treatment. In this work, we proposed a new model architecture that integrates DenseBlocks into the Inception Module that we refer to as DINET. The novel model captures multi-scale learned features while increasing parameter efficiency throughout the network. Our aim is to investigate the impact of maintaining a balanced tradeoff between model efficiency and effectiveness of preventing the problem of vanishing gradient. For our medical images classification problem, specifically using Chest x-ray images, experiment results using our DINET model shows modest performance improvements. In our future research directions, other versions of DINET modules will be experimented to shed more light on the impact of strategically positioning the DINET modules in the network to achieve the trade-offs of model efficiency and performance while increasing interaction of multi-scale learned features.

Key words: Chest X-rays, DCNN, Deep learning, DINET, healthcare, medical images.