1 of 1

Arushi Chauhan (2016019) Mehak Gupta(2016163) Raghav Sood(2016259)

Baseline Algorithm ( Self Implemented)

  1. External database at https://bit.ly/2FYd97D
  2. A.O. Voula, S.U. Akram, J.Kannala, ``Mask-RCNN And U-Net Ensembled For Nuclei Segmentation`` https://arxiv.org/pdf/1901.10170.pdf
  3. O. Ronneberger, P. Fischer, T. Brox, ``U-Net: Convolutional Networks for Biomedical Image Segmentation'' https://arxiv.org/abs/1505.04597
  4. K. He, G. Gkioxari, P. Dollar, R. Girshick, ``Mask R-CNN'' https://arxiv.org/abs/1703.06870
  5. Kaggle Data Science Bowl 2018 https://www.kaggle.com/c/data-science-bowl-2018

Results

References

Dataset

Inferences and Conclusion

Automating Nuclei Detection in Cells

We used the Kaggle Dataset for Data Science Bowl 2018 as our primary dataset. It consisted of 670 images and each image had 1 or more nuclei in it. Each image has a variable dimension. Some images have more than 350 nuclei in it. For each image, we have an ImageId and a set of masks each corresponding to a seperate nuclei which gives the exact set of pixels which belong to that nuclei. Masks are not allowed to overlap.

​

​

​

​

The total number of masks is 30,385. Along with the Kaggle dataset, we have also used another dataset which consists of Hematoxylin and Eosin (H&E) stained tissue images. Using (H&E) enhances the contrast between nuclei, epithelium, and stroma for examination under a microscope. In this dataset, there 30 tissue images. Overall 21,000 nuclear boundaries were annotated in this dataset.

To avoid overfitting we have heavily relied on dataset augmentation techniques as the given dataset is very small in size.

​

​

​

Our aim is to create an algorithm to automate nucleus detection. The computer model should be able to identify a range of cell nuclei across varied conditions like magnification, types of cells and image modality.

Problem Statement

Illnesses like cancer and heart diseases are responsible for 40% of all deaths worldwide. Nuclei identification is the starting point for analysing diseases since nuclei can be used to distinguish between different cells. Using this information, researchers can understand how cells react to a particular treatment, and thus, they can find effective cures for diseases. We intend to develop an algorithm to automate nuclei identification.

​

Introduction and Use Cases

An example of an image in the dataset and the corresponding masks

IIIT Delhi

Current Algorithm

The inferences we derived from the results are:

​

  1. False positive hurt the score much more than false negatives.
  2. Dilating and then eroding individual masks helps in getting better results.
  3. Model reaches the local minima much faster when we use ADAM optimizer instead of the default Stochastic Gradient Descent optimizer.
  4. Scaling the images in the preprocessing steps helps with the generalizability of the model.

​

Conclusion: Mask RCNN is a better model than UNet while performing instance segmentation when dataset is small.

The algoritihm we used is based on the Mask RCNN model. The steps can be summarized as follows:

​

We choose the U-Net model as a baseline for our instance segmentation task. Although the U-Net model has been implemented a number of times, we chose to implement it ourselves for our specific task. Before applying the U-Net model the following preprocessing steps were applied to the images:

  • Some of the masks contained multiple nuclei or nuclei with holes in it. These masks were manually fixed, to ensure each mask corresponds to only one nuclei.
  • The dataset was manually divided into 5 classes.
  • The dataset was split into 90% train and 10% validation sets using stratified split.
  • Images were converted from RGB to HSV space.

Along with preprocessing we also applied augmentation techniques like rotation, scaling, cropping,etc. to avoid overfitting. We applied the U-Net model as baseline along with the soft dice loss function.

With the advancement of techniques recently, a lot of research has been done in the field of automatic nuclei detection. Earlier, thresholding based techniques were used for detecting nuclei but they failed to detect clustered nuclei. An adaptive H-minima transform technique which adaptively suppresses the local minima was later proposed to solve this problem. Further speed ups were made by proposing a single pass based voting technique for nuclei detection followed by mean shift clustering to detect the nuclei centres. All these techniques have high computational complexity with a large number of cell nuclei.

​

Recently the problem of detecting cell nuclei has been treated as an instance segmentation problem. The two most popular techniques used for instance segmentation are: a) Mask RCNN [4] and b) UNet : a ConvNet for Biomedical Image Segmentation [3]. Mask RCNN is an extension of Faster RCNN which adds a branch for predicting an object mask in parallel with the existing branch for bounding box recognition. Mask R-CNN is simple to train and adds only a small overhead to Faster R-CNN, i.e. running at 5 fps. U-Net is a network which doesn't require a large number of training samples. It relies on the strong use of data augmentation to use the available annotated samples more efficiently.

We will use these two techniques in our approach to solve the problem

Literature Review

Baseline Algorithm

Current Algorithm

Fig1

IOU Score for UNet Model : 0.548

Fig2

Fig3

Fig4

Fig5

IOU Score for Mask RCNN : 0.6823

The current algorithm performed much better when compared to the baseline.

​

​

The evaluation metric we used is Intersection Over Union:

Template ID: bluecrosses Size: 36x48