Loading...

Histopathological Image Retrieval Using Self-Supervised Methods

Abbasi, Reza | 2024

0 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 57647 (19)
  4. University: Sharif University of Technology
  5. Department: Computer Engineering
  6. Advisor(s): Rohban, Mohammad Hossein
  7. Abstract:
  8. Identifying the type of disease from microscopic images taken from human tissue is considered a challenging task in cases where diagnosis is difficult and ambiguous for specialists. To resolve this ambiguity, pathologists spend a lot of time finding similar images in databases to determine the label of the ambiguous image based on other pathologists’ diagnoses of similar images. Therefore, automatic image retrieval is one of the active research topics in computer vision and medical image analysis. In this regard, various models have been introduced to enable doctors to perform this task in less time. These models work by receiving an image as input and returning the most biologically similar cases from the images in the database. Currently, there are two main problems in this field: The first is the lack of labeled medical data for training deep neural network models in this domain. The second problem is the low accuracy of many introduced models when faced with various lighting conditions, staining, and imaging in different cancers. In the first step of this project, we demonstrate through a series of experiments that the evaluation methods of previous works are not accurate and cannot be generalized to real-world conditions. In the next step, we design a new benchmark and use it to evaluate different models, showing that the models created so far for medical image retrieval have low accuracy under the new conditions we apply. Next, we propose self-supervised models as a solution to improve this low accuracy and show that in the experiments we have designed, these models perform better than supervised models. Finally, by introducing a new cost function, we fine-tune the models that were trained in a self-supervised manner and arrive at a model that outperforms all previously introduced models. To evaluate the proposed method, we use various datasets, including public datasets related to breast and colon cancer, and measure the accuracy of the proposed model in comparison with other existing models in this field. The datasets used in this project include CRC, BACH, and BRACS. The evaluation metrics used include recall rate, accuracy rate, and precision rate for the top k retrieved images
  9. Keywords:
  10. Image Retrieval ; Deep Learning ; Self-Supervised Learning ; Histopathology Images ; Disease Diagnosis

 Digital Object List