Summary:
Researchers at UCLA's Department of Radiological Sciences have developed an active learning methodology for digital pathology image analysis, addressing challenges posed by inconsistent annotations and revolutionizing the training of diagnostic algorithms.
Background:
The global digital pathology market is anticipated to experience significant growth in the coming years, with projections indicating an increase from $5.9 billion in 2022 to $10.7 billion by 2027. The World Health Organization (WHO) estimates a 70% rise in cancer diagnoses over the next two decades, alongside nearly 1.4 million misdiagnoses. Consequently, the expansion of digital pathology can be primarily attributed to the need for faster and more accurate diagnostic modalities for various chronic illnesses including cancer. Medical image analysis tasks have substantially benefited from machine learning techniques, particularly deep neural networks. However, achieving optimal performance necessitates accurate expert annotations, which can prove challenging and time-consuming. This is especially true for histopathology image analysis, where extensive professional training and expertise are required for proper annotation. Furthermore, there may be a lack of consensus among pathologists regarding annotations.
Numerous studies have explored approaches to address annotation difficulties in medical imaging. For instance, active learning (AL) has been employed to minimize the amount of labeled data required for learning. However, traditional AL methods do not address the issue of noisy labels. While other studies have attempted to identify and correct noisy labels in training data to enhance the performance of deep neural networks in medical image analysis, these methods do not differentiate between mislabeled and complex samples. Given the rising prevalence of cancer and the expanding elderly population, there is an urgent need to revolutionize the way image analysis algorithms are trained for pathology applications.
Innovation:
Professor Corey Arnold and his research team have developed a novel method for training pathology image analysis algorithms. Their approach allows for the automatic differentiation between mislabeled data in two categories of histopathological complexities: difficult-to-classify and easily-classified samples. As a result, the method leads to reduced time expenditure and enhanced accuracy
This invention from the University of California, Los Angeles (UCLA) advances active learning (AL) in medical image analysis by streamlining the training process. The novel methodology demonstrates promising results in prostate cancer Gleason grading, achieving a 40% reduction in the number of required annotations while maintaining performance levels comparable to fully supervised learning approaches.
Potential Applications:
• Prostate and other cancer diagnosis
• Point-of-care diagnosis
• Image analysis for general scientific research (electron microscopy images)
• General classification of object and texture for machine learning
Advantages:
• Improved prediction accuracy
• Requires less training data
• Distinguishes between noisy and complex data
• Detects and cleanses the noisy samples
• Fits for various histopathological image analysis scenarios
Development to Date:
First successful demonstration (first actual reduction to practice): Sep. 1st, 2020.
Patent Application:
System and method for pathology image analysis using a trained neural network and active learning framework
Related Papers:
Li, Wenyuan, Jiayun Li, Zichen Wang, Jennifer Polson, Anthony E. Sisk, Dipti P. Sajed, William Speier, and Corey W. Arnold. "PathAL: An Active Learning Framework for Histopathology Image Analysis." IEEE Transactions on Medical Imaging 41, no. 5 (2021): 1176-1187.
Reference:
UCLA Case No. 2021-134
Lead Inventor:
Prof. Corey Arnold; Dr. Wenyuan Li; Mr. Jiayun Li; Mr. William Speier.