A Multi-Task Pipeline with Specialized Streams for Classification and
Segmentation of Infection Manifestations in COVID-19 Scans
Abstract
We are concerned with the challenge of coronavirus disease (COVID-19)
detection in chest
X-ray and Computed Tomography (CT) scans, and the classification and
segmentation of related
infection manifestations. Even though it is arguably not an established
diagnostic tool, using machine
learning-based analysis of COVID-19 medical scans has shown the
potential to provide a preliminary
digital second opinion. This can help in managing the current pandemic,
and thus has been attracting
significant research attention. In this research, we propose a
multi-task pipeline that takes advantage
of the growing advances in deep neural network models. In the first
stage, we fine-tuned an
Inception-v3 deep model for COVID-19 recognition using multi-modal
learning, i.e., using X-ray and
CT scans. In addition to outperforming other deep models on the same
task in the recent literature,
with an attained accuracy of 99.4%, we also present comparative
analysis for multi-modal learning
against learning from X-ray scans alone. The second and the third stages
of the proposed pipeline
complement one another in dealing with different types of infection
manifestations. The former
features a convolutional neural network architecture for recognizing
three types of manifestations,
while the latter transfers learning from another knowledge domain,
namely, pulmonary nodule
segmentation in CT scans, to produce binary masks for segmenting the
regions corresponding to
these manifestations. Our proposed pipeline also features specialized
streams in which multiple deep
models are trained separately to segment specific types of infection
manifestations, and we show the
significant impact that this framework has on various performance
metrics. We evaluate the
proposed models on widely adopted datasets, and we demonstrate an
increase of approximately 4%
and 7% for dice coefficient and mean intersection-over-union (mIoU),
respectively, while achieving
60% reduction in computational time, compared to the recent
literature.