Cytology tests such as fine needle aspiration (FNA) and exclusion cytology test, play crucial role in some types of cancer treatment. Usually cancer is diagnosed by observations of cells in image-based cellular research. Because these processes are simple and repetitive, applying artificial intelligence (AI) makes the process much easier and efficiently. AI is also very powerful tools for cell segmentation which is important for detection and classification in pathology. Therefore, AI will be the key technology for segmentation of cancerous tissue which is the main process of the cancer diagnosis and treatment. To increase the reliability of our diagnostic model, we identified the distribution and number of cancer cells. For this purpose, we tried cell segmentation with different architectures based on Unet. After creating a mask, we used OpenCV’s Contour function to determine the number and distribution of cancer cells. The main goal is the segmentation of thyroid cancer data. Because our data does not have annotated images, we utilized similar data to develop a mask generation model of the data. Additionally, the thyroid cancer data is limited which belongs to our project and is not publicly available. We replace it with images from the paper “Medullary Thyroid Cancer” [6], which is the most similar data compare to ours.

A total of 2656 images are available. 1328 Original blood cell images with 1328 corresponding ground truths. Out of that, Jeet B Lahiri separated the training and testing sets with 1169 images and 159 images respectively. Data gernerate by Deponker, Sarker Depto, Shazidur Rahman, Md. Mekayel Hosen, Mst Shapna Akter, Tamanna Rahman Reme, Aimon Rahman, Hasib Zunai, M. Sohel Rahman and M.R.C.Mahdy

Distributed under the MIT License.

This dataset, also known as PanNuke, contains semi automatically generated nuclei instance segmentation and classification images with exhaustive nuclei labels across 19 different tissue types. The dataset consists of 481 visual fields, of which 312 are randomly sampled from more than 20K whole slide images at different magnifications, from multiple data sources. In total the dataset contains 205,343 labeled nuclei, each with an instance segmentation mask. Models trained on PanNuke can aid in whole slide image tissue type segmentation, and generalize to new tissues.

Distributed under the CC BY NC SA 4.0

Unet, Res_Unet, Attention_Unet, VGG11 & 16 Unet structures were used. Unet and Res_Unet were overfitting. We compared three models, Attention_Unet, VGG11 & 16 Unet. Based on IoU matrix, VGG16_RGB showed the best performance. Based on the effectiveness in removing impurities, lines and fluids, and good cell separation, Attention_Unet_Grayscael showed the best performance. Furthermore, we compared it with the Meta SAM model for 1000 images. We organized the run time and memory usage, and found that the performance was gotten marginally better. SAM uses Vit structure, which requires a large capitalization compared to based_Unet.

Based on the effectiveness in removing impurities, lines and fluids, and good cell separation, Attention_Unet_Grayscael performed the best.

We further compared it with the Meta SAM model for 1000 images. The performance was marginally better. We compared run time and memory usage.

SAM uses Vit structure, which requires a large capitalization compared to based_Unet.

[1] JEET B. LAHIRI. "Blood Cell Segmentation Dataset." Kaggle. 2023. https://www.kaggle.com/datasets/jeetblahiri/bccd-dataset-with-mask

[2] Olaf Ronneberger, Philipp Fischer, Thomas Brox. U-Net: Convolutional Networks for Biomedical Image Segmentation. arXiv, 2015. https://arxiv.org/abs/1505.04597v1

[3] Estibaliz Gómez de Mariscal, Martin Maška, Anna Kotrbová, Vendula Pospichalova, Pavel Matula & Arrate Muñoz-Barrutia. Deep-Learning-Based Segmentation of Small Extracellular Vesicles in Transmission Electron Microscopy Images. Scientific Reports, 2019. https://www.researchgate.net/publication/335802605_Deep-Learning-Based_Segmentation_of_Small_Extracellular_Vesicles_in_Transmission_Electron_Microscopy_Images

[4] Ozan Oktay, Jo Schlemper, Loic Le Folgoc, Matthew Lee, Mattias Heinrich, Kazunari Misawa, Kensaku Mori, Steven McDonagh, Nils Y Hammerla, Bernhard Kainz, Ben Glocker, Daniel Rueckert. Attention U-Net: Learning Where to Look for the Pancreas. arXiv, 2018. https://arxiv.org/abs/1804.03999

[5] Vladimir Iglovikov, Alexey Shvets. TernausNet: U-Net with VGG11 Encoder Pre-Trained on ImageNet for Image Segmentation. arXiv, 2018. https://arxiv.org/abs/1801.05746

[6] Michiya Nishino, Marc Philippe Pusztaszeri, Martha B. Pitman. Medullary Thyroid Carcinoma, In book: The Bethesda System for Reporting Thyroid Cytopathology (pp.157-175). ResearchGate, 2018. https://www.researchgate.net/publication/319449480_Medullary_Thyroid_Carcinoma

[7] Larxel. "Cancer Instance Segmentation and Classification 1,2,3" Kaggle. 2022. https://www.kaggle.com/datasets/andrewmvd/cancer-inst-segmentation-and-classification

[8] Gamper, Jevgenij and Koohbanani, Navid Alemi and Graham, Simon and Jahanifar, Mostafa and Benet, Ksenija and Khurram, Syed Ali and Azam, Ayesha and Hewitt, Katherine and Rajpoot, Nasir. PanNuke Dataset Extension, Insights and Baselines. arXiv, 2020. https://arxiv.org/abs/2003.10778

Distributed under the MIT License.