Authors:
Obioma Pelka
1
;
Felix Nensa
2
and
Christoph M. Friedrich
3
Affiliations:
1
University of Applied Sciences and Arts Dortmund and University of Duisburg-Essen, Germany
;
2
University Hospital Essen, Germany
;
3
University of Applied Sciences and Arts Dortmund, Germany
Keyword(s):
Biomedical Imaging, Deep Learning, Keyword Generation, Machine Learning, Multi-modal Representation, Transfer Learning, Radiographs.
Abstract:
As the number of digital medical images taken daily rapidly increases, manual annotation is impractical,
time-consuming and prone to errors. Hence, there is need to create systems that automatically classify and
annotate medical images. The aim of this presented work is to utilize Transfer Learning to generate image
keywords, which are substituted as text representation for medical image classification and retrieval tasks.
Text preprocessing methods such as detection and removal of compound figure delimiters, stop-words, special
characters and word stemming are applied before training the keyword generation model. All images are
visually represented using Convolutional Neural Networks (CNN) and the Long Short-Term Memory (LSTM)
based Recurrent Neural Network (RNN) Show-and-Tell model is adopted for keyword generation. To improve
model performance, a second training phase is initiated, where parameters are fine-tuned using the pre-trained
deep learning network Inception-ResNet-V2. For
the image classification tasks, Random Forest models trained
with Bag-of-Keypoints visual representations were adopted. Classification prediction accuracy was higher for
all classification schemes and on two distinct radiology image datasets using the proposed approach.
(More)