Augmented Mitotic Cell Count using Field Of Interest Proposal

Augmented Mitotic Cell Count using Field Of Interest Proposal

Marc Aubreville, Christof A. Bertram, Robert Klopfleisch , Andreas Maier
Abstract.

Histopathological prognostication of neoplasia including most tumor grading systems are based upon a number of criteria. Probably the most important is the number of mitotic figures which are most commonly determined as the mitotic count (MC), i.e. number of mitotic figures within 10 consecutive high power fields. Often the area with the highest mitotic activity is to be selected for the MC. However, since mitotic activity is not known in advance, an arbitrary choice of this region is considered one important cause for high variability in the prognostication and grading.
In this work, we present an algorithmic approach that first calculates a mitotic cell map based upon a deep convolutional network. This map is in a second step used to construct a mitotic activity estimate. Lastly, we select the image segment representing the size of ten high power fields with the overall highest mitotic activity as a region proposal for an expert MC determination. We evaluate the approach using a dataset of 32 completely annotated whole slide images, where 22 were used for training of the network and 10 for test. We find a correlation of r=0.936 in mitotic count estimate.

Augmented Mitotic Cell Count using Field Of Interest Proposal

Marc Aubreville, Christof A. Bertram, Robert Klopfleisch , Andreas Maier

Pattern Recognition Lab, Computer Sciences, Friedrich-Alexander-Universität Erlangen-Nürnberg

Institute of Veterinary Pathology, Freie Universität Berlin, Germany


marc.aubreville@fau.de


1 Introduction

One important aspect of tumor prognostication in human and veterinary pathology is the proliferative rate of the tumor cells, which is assumed to be correlated with the density of cells undergoing divison (mitotic figures) in a histology slide[1] and is applied as a criterion in almost all current tumor grading systems. However, mitotic activity is known to have large inter-observer variances [2], which consequentially strongly affects the histological grade assigned. One reason might be that the classification between mitotic and non-mitocic cells is not clearly defined and varies across labs, schools and even individuals [3, 4]. Another important reason for this is, that the distribution of mitotic cells in the slide is usually sparse with local changes in density across the specimen. In clinical practice, this sampling problem is dealt with by counting mitotic figures in ten fields of view at a magnification of 400 (high power fields, HPF), resulting in the mitotic count (MC). However, as shown previously [5], especially for low to borderline mitotic counts, semi-random selection of those ten high power fields is not sufficient for a reproducible MC determination. While examining larger areas would improve on this, it is not the method of choice given limited time budgets in pathology labs. As of this writing, completely algorithmic approaches for mitotic activity estimation lack the sensitivity and specificity that would be required to achieve clinical applicability. Further, purely algorithmic outcomes may be subject to hesitation from the pathology side, since automatic solutions that are not easily comprehensible for the medical expert, such as deep learning networks, may not be robust.

In this paper, we present an algorithmic approach that proposes a region of the area of 10 high power fields that is assumed to have the highest mitotic count within the slide. This has two positive aspects: While we still rely on the expertise of a pathologist to assess the actual mitotic activity, we limit the focus area to a defined field of interest in the image. Further, as this algorithmic answer will always be equal for the same image, it will allow us to differentiate the true inter-observer variance in an optimal setting when the area on the slide is already fixed. This region proposal will serve as an augmentation to the pathology expert.

2 Material

For this work, we annotated 32 histology slides of canine cutaneous mast cell tumors dyed with standard hematoxylin and eosin stain. The slides were digitized using a linear scanner (Aperio ScanScope CS2, Leica Biosystems, Germany) at a magnification of (resolution: 0.25 ). Contrary to popular other publicly available mitosis data sets, we did not pre-crop the whole slide images (WSI) but include all parts of the slide, including borders, which we consider important for a general applicability of the framework. All slides have been annotated by two pathologists using the open source annotation software SlideRunner [6]. Out of all cells annotated as mitotic figure, we only use those where both observers agreed upon being a mitotic figure. We arbitrarily chose 10 slides to be the test set, and 22 to be used for the training process. The data set includes slides of low, medium and high mitotic activity in both training and test set. In total, 45,811 mitotic cells have been annotated on all slides. To the best of our knowledge, this data set is unprecedented in size for any mitotic cell task and may serve as basis for many algorithmic improvements to the field.

3 Methods

A significant number of algorithmic approaches for mitosis detection have emerged very recently, most based on deep convolutional networks [7, 8, 9], making use of transfer learning [10] and hard-negative example mining [8, 7]. Typically, these algorithms use a two-stage approach, where in the first stage multiple regions of interest are detected and in a second stage classification is done according to being a mitotic figure or not. However, as also stated in the TUPAC challenge, automated identification of mitoses is only an intermediate step in tumor grading. F1-scores of up to 0.652 [9] have been achieved on the TUPAC challenge test data set. Current results are unlikely to reach clinical standards. Additionally, fully automated grading algorithms could run into acceptance problems, because robustness in a clinical workflow has yet to be proven. We regard mitosis detection as an intermediate step needed to propose a region of interest that could either be representing the statistics of the complete slide, or, as typically intended, represents the region of highest mitotic activity. For this approach, however, it is valid to not consider the object detection task of mitotic figures, but rather to derive maps where mitotic cells are located.

3.1 Mitosis as Segmentation Task

For the purpose of field of interest proposal, we consider mitotic figure detection a segmentation task, with mitotic figures being represented by filled circles. This enables the use of concepts like the dice coefficient (intersection/union) for both, evaluation as well as for optimization.

\hb@xt@

a) Relation between ground truth mitotic count (MC) prediction and estimated MC prediction on test set (r=0.936) b) MC distribution on test slides (ground truth). Red dashed line marks ground truth MC for proposed position.

4 Discussion

The mitotic figure prediction network scored in the same order as other algorithms on other data sets that also do mitosis detection. However, while the general problem of automatically identifying mitotic figures in WSI with sufficient accuracy for clinical application remains a challenge, the outcomes of these approaches might indeed serve as a surrogate for field of interest proposal and thus as a augmentation to the pathology expert. In future studies, it will have to be proven that clinical application of such augmentation methods will be able to reduce variability in MC determination.

References

  • [1] Elston CW, Ellis IO. Pathological prognostic factors in breast cancer. I. The value of histological grade in breast cancer: experience from a large study with long-term follow-up. Histopathology. 1991;19(5):403–410.
  • [2] Boiesen P, Bendahl PO, Anagnostaki L, Domanski H, Holm E, Idvall I, et al. Histologic grading in breast cancer: reproducibility between seven pathologic departments. Acta Oncol. 2000;39(1):41–45.
  • [3] Meuten DJ, Moore FM, George JW. Mitotic Count and the Field of View Area. Vet Pathol. 2016 Jan;53(1):7–9.
  • [4] Bertram CA, Gurtner C, Dettwiler M, Kershaw O, Dietert K, Pieper L, et al. Validation of Digital Microscopy Compared With Light Microscopy for the Diagnosis of Canine Cutaneous Tumors. Vet Pathol. 2018;55(4):490–500.
  • [5] Bonert M, Tate AJ. Mitotic counts in breast cancer should be standardized with a uniform sample area. BioMedical Engineering OnLine. 2017 Jan; p. 1–8.
  • [6] Aubreville M, Bertram CA, Klopfleisch R, Maier A. SlideRunner - A Tool for Massive Cell Annotations in Whole Slide Images. In: Bildverarbeitung für die Medizin 2018. Springer; 2018. p. 309–314.
  • [7] Cireşan DC, Giusti A, Gambardella LM, Schmidhuber J. Mitosis detection in breast cancer histology images with deep neural networks. MICCAI. 2013;16(Pt 2):411–418.
  • [8] Akram SU, Qaiser T, Graham S, Kannala J, Heikkilä J, Rajpoot N. Leveraging Unlabeled Whole-Slide-Images for Mitosis Detection. In: Computational Pathology and Ophthalmic Medical Image Analysis. Springer; 2018. p. 69–77.
  • [9] Paeng K, Hwang S, Park S, Kim M. A unified framework for tumor proliferation score prediction in breast histopathology. In: Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support. Springer; 2017. p. 231–239.
  • [10] Li C, Wang X, Liu W, Latecki LJ. DeepMitosis: Mitosis detection via deep detection, verification and segmentation networks. Med Image Anal. 2018 Apr;45:121–133.
  • [11] Otsu N. A Threshold Selection Method from Gray-Level Histograms. IEEE Trans Syst Man Cybern. 1979;9(1):62–66.
  • [12] Ronneberger O, Fischer P, Brox T. U-Net - Convolutional Networks for Biomedical Image Segmentation. MICCAI. 2015;9351(Chapter 28):234–241.
  • [13] Rahman MA, Wang Y. Optimizing intersection-over-union in deep neural networks for image segmentation. In: International Symposium on Visual Computing. Springer; 2016. p. 234–244.
Comments 0
Request Comment
You are adding the first comment!
How to quickly get a good reply:
  • Give credit where it’s due by listing out the positive aspects of a paper before getting into which changes should be made.
  • Be specific in your critique, and provide supporting evidence with appropriate references to substantiate general statements.
  • Your comment should inspire ideas to flow and help the author improves the paper.

The better we are at sharing our knowledge with each other, the faster we move forward.
""
The feedback must be of minimum 40 characters and the title a minimum of 5 characters
   
Add comment
Cancel
Loading ...
297531
This is a comment super asjknd jkasnjk adsnkj
Upvote
Downvote
""
The feedback must be of minumum 40 characters
The feedback must be of minumum 40 characters
Submit
Cancel

You are asking your first question!
How to quickly get a good answer:
  • Keep your question short and to the point
  • Check for grammar or spelling errors.
  • Phrase it like a question
Test
Test description