A digital pathology platform for automated diagnosis and classification of cervical cancer from pap-smear images (digpath)
![Thumbnail](/xmlui/bitstream/handle/123456789/351/A%20digital%20Pathology%20Platform%20for%20Automated%20Diagnosis%20and%20Classification%20of%20Cervical%20Cancer%20from%20pap.pdf.jpg?sequence=4&isAllowed=y)
View/ Open
Date
2019Author
Wasswa, William
Ware, Andrew
Basaza-Ejiri, Habinka Annabella
Obungoloch, Johnes
Metadata
Show full item recordAbstract
Background: Globally, cervical cancer ranks as the fourth most frequent cancer in women with
an estimated 570,000 new cases in 2018, representing 6.6% of all female cancers. Approximately
90% of deaths from cervical cancer occur in low and middle income countries with Uganda
ranked 7th among the countries with highest incidences of cervical cancer in Africa. However,
cervical cancer can be prevented through regular screening.
Objective: To develop a digital pathology platform for automated diagnosis and classification of
cervical cancer from pap-smear images.
Methods: The digital pathology platform consists of three entities: (1) An automated low-cost
digital microscope slide scanner producing quick, reliable and high-resolution cervical cell
images from pap-smears. It is 3D printed and costs less than 500USD compared to the
commercial-microscopes costing over 2000USD. It has an objective-lens from traditional
microscopes; a Logitech web-camera for image capture placed at a distance from the lens
calculated using the 4f principle, a motorized stage driven by 2 stepper motors for XYmovement and a third motor for focusing (Z-movement). Image capture is by a developed
software in c++. Auto focusing is by an algorithm based on Fast-Fourier transform. The stage
control is accomplished using the grbl library as used in the cnc machines to provide precise
movements. Limit switches are used for position sensing, (2) An automated pap-smear analysis
tool for diagnosis and classification of cervical cancer from pap-smear images. For image
analysis pipeline, scene segmentation is achieved using a sequential elimination approach, image
segmentation is achieved using trainable classifier, feature selection is with Simulated annealing
coupled with a wrapper filter and classification is based on an enhanced fuzzy c-means algorithm
and (3) Cervical cancer risk factors evaluation to automatically assess the likelihood of
contacting cervical cancer given the risk factors implemented using Mamdani fuzzy logic based
on the knowledge base provided by experts.
Results: The evaluation of the automated pap-smear analysis was carried out on three different
datasets (single cell images, multiple cell images and pap-smear slide images from a pathology
unit. Overall classification accuracy, sensitivity and specificity of ‘98.88%, 99.28% and
97.47%’, ‘97.64%, 98.08% and 97.16%’ and ‘95.00%, 100% and 90.00%’ were obtained for
each dataset respectively. The evaluation and testing conducted confirmed the rationale of the
proposed approach that the selection of good features embeds sufficient discriminatory
information that can increase the accuracy of cervical cancer classification. Evaluation of the
fuzzy inference rules for cervical cancer risk assessment showed that the diagnosis produced by
the tool were in agreement with the diagnosis from the cytopathologist. Evaluation of the
microscope slide scanner showed that the developed microscope can provide high-resolution
digital images. For a conventional pap-smear slide, the developed microscope can produce a
digital image in less than three minutes with resolutions of 1.10µm and 0.42µm using a 10x lens
and a 40x lens respectively. The image quality is comparable to high-end commercial
microscopes at a cost of less than $500.
Conclusion: The major contribution of this platform in cervical cancer screening is that it
reduces on the time required by the cytotechnician to screen very many slides by eliminating the
obvious normal ones, hence more time can be put on the suspicious slides. The proposed
platform has the capability of analyzing a full-slide in less than four minutes as opposed to 5-10
minutes per slide in the manual analysis.
Recommendations: Despite the high performance of the approach, it, however, uses a chain of
number of methods which makes it computationally heavy and this is a limitation of the
proposed method. In the future deep learning, approaches will be explored to reduce the
complexity of the approach for clinical use.
Collections
- Research Articles [25]