Dear Aswathy M a

These articles might help, have a look:

1. https://www.frontiersin.org/articles/10.3389/fbioe.2019.00053/full

2. Article Large scale tissue histopathology image classification, segm...

3. Article Segmentation and Classification in Digital Pathology for Gli...

4. Article Histopathology Image Segmentation and Classification for Can...

5. https://www.nature.com/articles/s41598-021-90444-8

6. https://www.degruyter.com/document/doi/10.1515/jisys-2018-0316/html

Kind Regards

Qamar Ul Islam

Hi,

Segmentation means you are trying to locate the ROI (region of interest) which should, at least theoretically, has a strong predictive power. A good number of prior research work consider images as they are by extracting blindly features in either the spatial or frequency domains or in both. Some researchers are concerned about shape descriptions (of tumors), thus the need for segmentation.

Now, back to your question:

"Can we classify histopathology images without segmentation?" Yes, we can, however, the classification accuracy and generalisability might be vulnerable.

It could be much meaningful to target the ROI and extract salient features (that must be invariant to scale, intensity -agent concentration-, rotation, etc.). If the segmentation is done robustly, this will even benefit deep learning (partial interpretability) as we did in the study of Wu and Cheddad (2019)

https://ieeexplore.ieee.org/document/8936078

Article Large scale tissue histopathology image classification, segm...

So if we do classification without segmentation, then how will the model extract nuclei features like area, perimeter etc.

Due to the whole-slide histopathology image’s high-resolution, one approach to achieve the classification of a whole-slide image is by dividing it into smaller image tiles and processing each image tile independently in parallel on a cluster of computer nodes. See, 5.3. Multi-scale feature extraction and Figure 5.7 in the below reference.

“So if we do classification without segmentation, then how will the model extract nuclei features like area, perimeter etc.” The model in this case extracts global features not labelled to any structure (cell nuclei, glands, lymphocytes). For example, a model can rely on features extracted by detecting cell clusters, it can detect textural features, it can detect structures based on colour transformation +/- clustering, it can use features extracted from 2D filtered images or from regions identified by template matching, etc. A useful section, in the review paper, is discussing some of these options around Table 5.1 (size and shape are just part of a large possible features) and Table 5.2.

So, if you are new to the domain, my advice is to depart from the below review paper to give you a clear overview of the different options you have and the common procedures in this area. The paper is quite old (2009) but since you are not interested in deep learning that should not matter.

======

Gurcan, et al. (2009) “Histopathological Image Analysis: A Review”

https://ieeexplore.ieee.org/document/5299287 https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2910932/