Download as docx, pdf, or txt
Download as docx, pdf, or txt
You are on page 1of 3

THE COMPLETE WORKING OF OUR

CANCER DETECTION PROJECT


Project overview:
We have built a cancer detection model for detecting colon cancer with
the help of high quality histopathology images. In this model we have
incorporated the useful effects of various deep learning techniques like YOLO,
CNN and a lot more. So hereafter I will explain you about my project, the
different approaches that we have used, how they work, what happening at
the back and all the thing one needs to know about our project.
To start off with, in this specific model we have used two critical
approaches which are:
 A cell detector approach and
 The classifier approach

Describing about the cell detector approach it is basically a YOLO model


that is capable of detecting the parts where a cancer cell is present in a WSI
image. The process commences by dividing the complete WSI into smaller,
manageable sections, or slices. This step ensures that our model can efficiently
process and analyze the entire image. Our YOLO model is then applied to each
sliced image. The model conducts inference to determine the presence of
cancerous and benign cells. It generates probability scores for the identified
regions, allowing for differentiation between the two cell types. Following the
inference process, our model tabulates the results. It counts the number of
cancerous and benign cells detected throughout the entire WSI image. This
information is presented in a table,
including percentage estimates based
on our model's predictions.
Figure 1 Predictions made on the image.

The classifier approach uses a modified InceptionV3 model. WSIs are


initially fed into the model and split into tiles of size 5000x5000. These tiles are
stored in a designated folder for further processing. To isolate tissue regions, a
random forest algorithm is employed. This step ensures that the model focuses
only on areas containing relevant information. Inference is performed using a
sliding window approach, where the model analyzes each window individually.
The output assigns each window to one of the eight categories: Tumor,
Stroma, Complex, Debris, Mucosa, Adipose, or Empty. Each category is
associated with a specific color. The model uses this color scheme to generate
a new image, providing a visual representation of the different tissue classes
within the image. Heat maps are produced for each category, reflecting the
distribution of that class within the tissue. Merging these heat maps results in
a visual representation of the predominant tissue class.

Figure 1Above Figure shows the prediction for the given Tissue image as a tumor case.
Figure 3 Flow diagram of Model working.

You might also like