Article Text

Download PDFPDF

28 Deep learning based cytology negative sample screening method
  1. T Konishi1,
  2. J Kolodziejczyk2,
  3. K Saito3,
  4. H Hayashi4,
  5. T Iwabuchi5,
  6. R Nakamura6,
  7. Y Akagami7,
  8. K Washiya8,
  9. K Abe9 and
  10. H Nanjo10
  1. 1Cellspect Co.- Ltd, Data Analytics, Morioka, Japan
  2. 2AI Okinawa, n/a, Chatan, Japan
  3. 3Cellspect Co.- Ltd, Business development, Morioka, Japan
  4. 4Cellspect Co.- Ltd, Clinical Chemistry Development development, Morioka, Japan
  5. 5Cellspect Co.- Ltd, n/a, Morioka, Japan
  6. 6Akita Industrial Technology Center, Advanced Processing Technology Development Section, Akita, Japan
  7. 7Akita Industrial Technology Center, Director-General, Akita, Japan
  8. 8Akita Karyology and Histology Research Center, Research and Development Department, Akita, Japan
  9. 9Akita Karyology and Histology Research Center, President-director, Akita, Japan
  10. 10Akita University Hospital, Division of Clinical Pathology, Akita, Japan


Objectives We show that deep learning can be used to create high accuracy cells classifiers that can support cytologists in their work.

Two deep learning models are presented. The first model is used for cervical cancer cytology negative/positive prescreening and we show its high specificity at 100% Negative Predict Value (NPV). The second model is used to predict among 15 categories common to cervical analysis, as listed in figure 1(b).

Our classifiers include Grad-Cam visualizations that show both models concentrate on relevant areas of images.

Abstract 28 Figure 1

(a) Confusion matrix of prescreening model; (b) Accuracy of 15 categories classification model; (c) Grad-Cam visualizations

Methods Classifiers are based on deep convolutional networks. The crucial aspects in achieving high accuracy are data augmentation and focal loss.

Grad-Cam visualizations are used to explain models’ reasoning.

Results Prescreening model, based on 2700 samples, achieves 95.89% overall accuracy and 91.62% specificity at 100% NPV (figure 1(a)) ; second stage screening model, based on approximately 9500 samples, achieves 92.03% overall accuracy (figure 1(b)). Grad-Cam visualizations show that both models concentrate on relevant areas of the image when making predictions (figure 1(c)).

Conclusions We show that deep learning based classifiers can be useful in supporting cytologist in their work. Approximately 92% of samples that cytologist screen for cervical cancer are negative. With our model cytologists can concentrate their efforts on only positive samples and a small number of false positives predictions.

Statistics from

Request Permissions

If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.