|
|
An Automatic Liver Tumor Detection Method Using Moving Means
Laramie Paxton, Marian University, USA
Yufeng Cao, University of Southern California, USA
Kevin Vixie, Washington State University, USA
Yuan Wang, Washington State University, USA
Chaan Ng, University of Texas, USA
Brian Hobbs, Cleveland Clinic, USA
Abstract:
We present an automatic liver segmentation method that utilizes the
time series data in conjunction with the Boykov-Kolmogorov (BK) graph
cut algorithm and uses a novel approach of moving means for each of
the sample healthy and tumor tissue intensities (from a separate data
set) to iterate and improve the initial graph cut segmentation. Thus,
there is no training process required since the initial sample means are
computed in advance using Regions of Interest provided by radiologists.
This method provides a reasonable degree of accuracy for an automatic
segmentation scheme, yielding a mean Dice similarity coefficient (DSC)
of 77 percent, a relative volume di erence (RVD) of 21.6 percent, and a
volumetric overlap error (VOE) of 35.7 percent. The algorithm is simple
to implement computationally, and the mean runtime of 5.1 minutes is
reasonable given that no training process is necessary.
The main contribution of this model is to allow the healthy and tumor means to move
so that a more optimal segmentation can be obtained.
Full Paper (in PDF)
|
A Comparison of Feature Vectors in a Graph Cut-Based Liver Segmentation Algorithm
Laramie Paxton, Marian University-Wisconsin, USA
Yufeng Cao, University of Southern California, USA
Kevin Vixie, Washington State University, USA
Yuan Wang, Washington State University, USA
Chaan Ng, University of Texas, USA
Brian Hobbs, Cleveland Clinic, USA
Abstract:
Liver image segmentation presents a challenging set of conditions and is
an active area of research. In this paper, we compare the effectiveness of
five different feature vectors used in a preprocessing step for a graph cutbased
semi-automatic liver segmentation algorithm. The feature vectors
tested are formed using a median filter, averaging filter, Gaussian filter,
neighborhood, and novel use of time series data. When compared to
the expert-provided ground truth, the time series approach outperforms
the others and yields results comparable to other recent models in the
literature, giving a mean volume error (VOE) of 32.9 percent, mean Dice
similarity coefficient (DSC) of 0.8, and mean runtime of 74 seconds. We
also include a modified boundary term in the energy functional and normalize
both terms in order to avoid further scaling of the boundary term.
In place of a training process, we utilize sample Regions of Interest provided
by expert radiologists to compute sample vector means for healthy
and tumor tissues that are used in the regional term of the functional.
Contribution: The time series feature vector method represents a novel
approach that utilizes the time series data obtained from a sequence
of 59 CT scans as a preprocessing step, along with using a simplified
boundary term in the energy functional.
Full Paper (in PDF)
|
Parkinson's Disease Detection Using ResNet50 with Transfer Learning
Nusrat Jahan, Jagannath University, Bangladesh
Arifatun Nesa, Jagannath University, Bangladesh
Md. Abu Layek, Jagannath University, Bangladesh
Abstract:
Parkinson's disease (PD) is an incurable neurological disorder disease.
But there is still no standard medical provision to identify Parkinson's
disease. In this study, a fine motor symptom that is sketching has been
studied. The experiments are done on a significant number of PD patients
and Healthy Group (without PD). We proposed a system that can
determine the sketching and reports whether a PD patient's sketch or
not. Deep learning algorithms can deal with the solution of different
brain generalizing neural networks with the same design. Thus, we applied
Convolutional Neural Network (CNN) to classify sketched images
to discriminate or identify Parkinson's Disease (PD) affected patients
from the regular healthy (without PD) control group. The experiment
was done on different CNN models with transfer learning method and applying
on Spiral and Wave sketched data. The proposed system achieved
96.67% accuracy on the ResNet50 model with spiral sketching.
Contribution: The main contribution of this paper is,
we have used Transfer learning which enhanced the model performance.
Full Paper (in PDF)
|
Bangla-English Neural Machine Translation with
Bidirectional Long Short-Term Memory and Back
Translation
Arna Roy, Khulna University of Engineering & Technology, Bangladesh
Argha Chandra Dhar, Khulna University of Engineering & Technology, Bangladesh
M. A. H. Akhand, Khulna University of Engineering & Technology, Bangladesh
Md Abdus Samad Kamal, Gunma University, Japan
Abstract:
Machine translation (MT) has recently drawn attention to the automatic
translation of the text, documents, or webpages from one language to
another. Among various MT approaches, neural MT (NMT) is the most
feasible method, a data-driven approach consisting of special neural networks.
Among thousands of natural languages, remarkable efforts on MT
are concentrated on a few languages only; and the research is very
limited for many major languages such as Bangla. The study aims to build
an effective NMT system for Bangla-English MT. Bidirectional Long
Short-Term Memory (BiLSTM), a popular deep learning method for sequential
data operation, is considered in the present study. Attention
mechanism with the BiLSTM model and a special data augmentation
mechanism, called Back Translation (BT), are the significant features of
the proposed model. The proposed model outperforms the prominent
models for Bangla to English MT while tested on a benchmark dataset.
Contribution: A BiLSTM with attention mechanism
is proposed that is trained considering BT and found effective
for lowresource Bangla-English MT cases.
Full Paper (in PDF)
|
Comparison of Prognostic Determinants after Myocardial Infarction using Holter ECG Data at 72-h
Emi Yuda, Tohoku University, Japan
Itaru Kaneko, Tohoku University, Japan
Yutaka Yoshida, Nagoya City University, Japan
Junichiro Hayano, Heartbeat Science Lab Co., Ltd., Japan
Abstract:
Development of in medical sensor technology, it is
possible to measure human bio-signals over long periods of time.
In particular, electrocardiogram (ECG) data obtained by long-term
measurement in hospitals can contribute to the construction
of bioindicators for human diseases, which have high prognostic
power for cardiac diseases. In previous studies, it has predicted
the presence of myocardial necrosis, vascular occlusion, and
myocardial ischemia mainly by detecting characteristic ECG
findings such as abnormal Q waves, ST interval elevation, and
coronary T waves from ECG waveform. In this study, we
compared heart rate variability (HRV) indices predictive of
myocardial infarction calculated from 72-hour Holter ECG RR
interval data with indices calculated from 24-hour data. The
HRV indices of 5 subjects in the young group (mean 22 y) and 5
subjects in the middle-aged group (mean 46 y) were compared,
and we revealed the usefulness of the 72-hour data for some
indices, such as standard deviation NN interval (SDNN).
Contribution: We have shown that it is desirable to analyze data
obtained from long-term measurements in order to calculate
prognostic indices for cardiac diseases using heart HRV indices.
Full Paper (in PDF)
|
|
