Development of a software package for high/low risk cancer stratification using machine learning
Project goal: In this project, we propose to develop a software package for the quantitative analysis of immunohistological images using machine learning. The developed tool will make it possible to conduct a more accurate and standardized diagnosis of IHC samples, excluding factors such as bias and subjective assessments of specialists. Also, the program can potentially be used in monitoring the course of the disease and testing the effectiveness of therapy.
Project description: Immunohistochemical examination (IHC) is a tissue analysis method used to diagnose various diseases, including cancers, when standard histology is insufficient or when specific molecular tumor parameters need clarification. It helps identify tumor types by detecting specific proteins, assessing treatment effectiveness, and determining drug sensitivity or hereditary disease predisposition. Routine histology often fails to provide a final diagnosis or identify tumor markers for early cancer detection and treatment selection. IHC results are typically interpreted qualitatively and subjectively, though quantitative assessment provides more accurate marker percentages. Factors like paraffin block quality and pathologist expertise affect result accuracy, as IHC lacks full standardization. Artificial intelligence (AI) enhances IHC by enabling fast, accurate, and standardized analysis of protein expression in tissues. In Kazakhstan, quantitative IHC analysis is not yet used diagnostically, but its implementation could improve diagnostic accuracy and treatment strategies
Project facilitators:
Аршат Уразбаев - PI
Асхат Мынбай
Елдар Байкен
Замарт Рамазанова
Realisation period: 2023-2025
Expected results: We will develop a software package using machine learning for quantitative analysis of immunohistological images, enabling accurate and standardized diagnosis by minimizing bias and subjective pathologist assessments. The tool will monitor disease progression, evaluate therapy effectiveness, and determine optimal therapeutic strategies for patients. We will identify the most efficient machine learning method to analyze IHC samples and train the AI to accurately detect stained areas based on pathologist interpretations. The program will undergo preliminary quantitative analysis of IHC samples and build predictive models. It will be tested with pathologists, therapists, and diagnostic centers, comparing results with clinical parameters like blood/urine tests and expert conclusions. This approach will enhance the program’s accuracy and efficiency, with results informing personalized therapy strategies. Project progress will be published in Scopus-indexed Q1/Q2/Q3 scientific journals
Methodology: The project methodology was implemented by developing a software package for quantitative analysis of immunohistochemical (IHC) images using machine learning, based on the laboratory's experience in tumor segmentation on lung CT. IHC samples were collected from patients with their informed consent at the "Hospital of the Medical Center of the Presidential Administration of the Republic of Kazakhstan", using only high-quality images. The analysis process was divided into three stages: tissue type classification using persistent diagrams to determine topological characteristics, image segmentation using adapted methods such as STARDIST and our own algorithm based on intensity gradient, and cell classification by color and shape using a neural network. Various preprocessing methods were applied for segmentation depending on the tissue type, and cell classification was based on labeled data divided into categories (labeled, unlabeled, false). A neural network with a 100x100 input matrix and three output neurons was trained on two-thirds of the data, and a third was used for control. The segmentation and classification results were tested for cell counting and diagnosis.
Contacts: arshat@gmail.com
Project goal: In this project, we propose to develop a software package for the quantitative analysis of immunohistological images using machine learning. The developed tool will make it possible to conduct a more accurate and standardized diagnosis of IHC samples, excluding factors such as bias and subjective assessments of specialists. Also, the program can potentially be used in monitoring the course of the disease and testing the effectiveness of therapy.
Project description: Immunohistochemical examination (IHC) is a tissue analysis method used to diagnose various diseases, including cancers, when standard histology is insufficient or when specific molecular tumor parameters need clarification. It helps identify tumor types by detecting specific proteins, assessing treatment effectiveness, and determining drug sensitivity or hereditary disease predisposition. Routine histology often fails to provide a final diagnosis or identify tumor markers for early cancer detection and treatment selection. IHC results are typically interpreted qualitatively and subjectively, though quantitative assessment provides more accurate marker percentages. Factors like paraffin block quality and pathologist expertise affect result accuracy, as IHC lacks full standardization. Artificial intelligence (AI) enhances IHC by enabling fast, accurate, and standardized analysis of protein expression in tissues. In Kazakhstan, quantitative IHC analysis is not yet used diagnostically, but its implementation could improve diagnostic accuracy and treatment strategies
Project facilitators:
Аршат Уразбаев - PI
Асхат Мынбай
Елдар Байкен
Замарт Рамазанова
Realisation period: 2023-2025
Expected results: We will develop a software package using machine learning for quantitative analysis of immunohistological images, enabling accurate and standardized diagnosis by minimizing bias and subjective pathologist assessments. The tool will monitor disease progression, evaluate therapy effectiveness, and determine optimal therapeutic strategies for patients. We will identify the most efficient machine learning method to analyze IHC samples and train the AI to accurately detect stained areas based on pathologist interpretations. The program will undergo preliminary quantitative analysis of IHC samples and build predictive models. It will be tested with pathologists, therapists, and diagnostic centers, comparing results with clinical parameters like blood/urine tests and expert conclusions. This approach will enhance the program’s accuracy and efficiency, with results informing personalized therapy strategies. Project progress will be published in Scopus-indexed Q1/Q2/Q3 scientific journals
Methodology: The project methodology was implemented by developing a software package for quantitative analysis of immunohistochemical (IHC) images using machine learning, based on the laboratory's experience in tumor segmentation on lung CT. IHC samples were collected from patients with their informed consent at the "Hospital of the Medical Center of the Presidential Administration of the Republic of Kazakhstan", using only high-quality images. The analysis process was divided into three stages: tissue type classification using persistent diagrams to determine topological characteristics, image segmentation using adapted methods such as STARDIST and our own algorithm based on intensity gradient, and cell classification by color and shape using a neural network. Various preprocessing methods were applied for segmentation depending on the tissue type, and cell classification was based on labeled data divided into categories (labeled, unlabeled, false). A neural network with a 100x100 input matrix and three output neurons was trained on two-thirds of the data, and a third was used for control. The segmentation and classification results were tested for cell counting and diagnosis.
Contacts: arshat@gmail.com