AI/ML is being explored to improve retinoblastoma (RB) detection, the most common childhood eye cancer. Building on previous work with an Asian Indian cohort, researchers tested an AI model’s ability to detect and classify RB in a multiracial group. Despite varying racial representation, the model was retrained and achieved impressive results: 97% accuracy for RB detection and high accuracy (93-99%) in classifying tumor groups (A-E) according to the International Classification of Retinoblastoma.
The study titled “Artificial Intelligence and Machine Learning in Ocular Oncology, Retinoblastoma (ArMOR): Experience with a Multiracial Cohort” published in “Cancers” (2024) discusses the challenges related to the color variation in fundus images due to differences in melanin concentrations across different races, which can impact the accuracy of artificial intelligence and machine learning models in diagnosing retinoblastoma. It examines the performance of an AI model previously validated in an Asian-Indian cohort in a more diverse population.
The study was a collaborative project involving experts from India and the United States. Vijitha S. Vempuluru and Rajiv Viriyala contributed equally to the work, with Swathi Kaliki serving as the corresponding author. All three are affiliated with the Operation Eyesight Universal Institute for Eye Cancer at LV Prasad Eye Institute in Hyderabad, India. The team also included Patanjali Bhamidipati from the International Institute of Information Technology in Hyderabad, and Krishna Kishore Dhara from the Center for Innovation at Bourntec, also in Hyderabad, bringing expertise in AI and technology to the project. Clinical and ocular oncology expertise was provided by Virinchi Ayyagari and Komal Bakal from the Operation Eyesight Universal Institute, as well as renowned experts Sandor R. Ferenczy and Carol L. Shields from Wills Eye Hospital at Thomas Jefferson University in Philadelphia, USA. This diverse team combined their knowledge in eye cancer, artificial intelligence, and technology to improve early detection of retinoblastoma.
Here is the Summary of the Study:
Introduction to AI/ML in Retinoblastoma:
– Explores the current trends in artificial intelligence and machine learning applications for intraocular tumors.
– Highlights retinoblastoma as the most prevalent eye cancer in children, underscoring the importance of early detection.
Objective of the Study:
– Aims to develop a screening tool using AI/ML for the detection of retinoblastoma (RB).
– Focuses on enhancing an AI model previously trained on an Asian Indian cohort.
Methodology Overview:
– Utilizes a retrospective observational study design to analyze fundus images across different racial groups.
– Incorporates a total of 2473 images, categorizing eyes by race based on official classifications.
Demographic Profile of Participants:
– A total of 210 eyes were classified, with 73% from White, followed by 18% African American, and smaller percentages from other races.
– 7% of the images did not report any race, providing a diverse dataset for the study.
AI Model Training Results:
– The model was retrained specifically considering the racial differences affecting image color and melanin concentration.
– Achieved a sensitivity of 93% and specificity of 96% for detecting retinoblastoma across the multiracial cohort.
Classification Accuracy:
– The model demonstrated high accuracy in classifying tumors into the International Classification of Retinoblastoma groups A to E.
– Reported specific accuracies of 98%, 93%, >99%, 94%, and 93% for each respective group.
Implications for Future Research:
– Identifies potential improvements for AI models considering racial diversity in healthcare applications.
– Encourages further research into AI/ML capabilities in enhancing early detection and treatment outcomes for retinoblastoma.
Conclusion:
– Demonstrates the feasibility of AI in screening for retinal diseases, emphasizing its role in enhancing clinical methods.
– Suggests that AI can bridge disparities in healthcare outcomes by providing accurate diagnoses irrespective of racial variations.
Introduction to Retinoblastoma and AI:
– Retinoblastoma (RB) is a pediatric eye cancer where AI has significantly advanced detection methods since 2017.
– Early detection of RB is critical, particularly in low-income regions where monitoring treatments is less prioritized.
AI Model Performance by Race:
– Studies show that AI models trained on a single race have limited effectiveness in detecting RB in other racial groups.
– Retraining the AI model with diverse data improved its sensitivity and specificity for detecting RB across different races.
Significance of Data Collection:
– A comprehensive multi-racial image database was created to enhance the AI model’s accuracy in detecting RB.
– Data regarding patient race was meticulously gathered to understand variations in fundus pigmentation affecting detection.
Methodology of Feature Extraction:
– AI/ML models used for feature extraction included Mobile-Net v2 SSD and various computer vision techniques.
– Initial models trained only on Asian Indian data performed poorly on a diverse cohort, indicating the need for broader training.
Challenges with Fundus Imaging:
– Differences in fundus color significantly affected detection capabilities due to varying melanin levels among races.
– Smaller tumors (groups A and B) presented more challenges due to their blending with the fundus background.
Improvements in AI Model:
– The model was revised to improve optical feature detection, particularly for blood vessels and tumors.
– Color segmentation and contrast enhancements were introduced to accurately distinguish features across racial images.
Retraining and Model Accuracy:
– The AI model required extensive retraining due to racial differences in visual data, focusing on 14 key features.
– An XGBoost ML model was employed to effectively classify and group RB after improving feature extraction methods.
Patient-Level Diagnosis Methodology:
– A systematic aggregation process was developed to accurately assign group labels based on multi-image evaluations for each patient.
– Outlier removal was a key step in refining the group assignment process to ensure accurate diagnoses.
Image Processing Steps:
– An aggregation algorithm filters outliers from eye images based on frequency,
– Images are grouped by the most common label with a threshold of 65%.
Dataset Overview:
– The study used 2473 images from 210 eyes, classified across multiple ICRB groups.
– Group B had the highest representation, primarily amongst the White race.
Performance Metrics:
– The AI model achieved high sensitivity (93%) and specificity (96%) across all images.
– Performance metrics varied per ICRB group, showcasing distinct values for each classification.
Accuracy Metrics:
– Overall accuracy reached 94% for images and 97% for eyes when detecting RB.
– Misclassification rates were noted, with under-classification at 11% and over-classification at 5% for images.
Sensitivity Analysis:
– Sensitivity of the AI model was highest for Asian Indian cohorts at 100%.
– Lower sensitivity was observed for White (86%) and African American (89%) groups.
Impact of Race on AI Performance:
– Racial factors influenced the AI model’s accuracy, necessitating diverse training datasets.
– Background color variances in fundus images affect model performance in different ethnicities.
Need for Diverse Training Datasets:
– Ensuring high-quality and diverse datasets is essential for training effective AI models.
– Previous studies illustrate improved AI sensitivity and specificity with larger datasets.
Future Directions:
– Future models should focus on improving discrimination across varying racial backgrounds.
– Training AI on diverse cohorts can significantly enhance detection capabilities for RB.
Study Purpose:
– To improve the accuracy of an AI model for detecting small tumors in retinoblastoma (RB) across a multiracial cohort.
– Highlights the performance variation of the AI model based on racial differences, necessitating retraining.
Model Development:
– A single generic AI model was developed for broader deployment instead of multiple race-specific models.
– This versatility allows for efficient retraining and testing with multiracial data without requiring prior racial identification.
Improving Accuracy:
– Retraining the AI model on a diverse database led to enhanced accuracy in detecting smaller tumors.
– The study emphasizes the need for further training with a wider range of images from different races.
Limitations:
– The study’s images were predominantly from White patients, leading to potential biases in model performance for underrepresented racial groups.
– Uneven distribution of images across different cohorts limits the AI model’s generalizability beyond the studied races.
Future Directions:
– Recommendations for multicentric data-sharing studies to broaden the dataset, particularly for underrepresented racial groups.
– The need for improved methodology to ensure uniformly high-quality images across diverse cohorts is highlighted.
Screening Program Considerations:
– An open question remains whether screening programs based on this AI model should be race-specific or universal.
– The findings underline the importance of addressing race-specific challenges in RB detection.
Research Contributions:
– This study is among the first to explore AI’s capabilities in detecting and classifying RB in diverse populations.
– It establishes a foundation for future research collaborations aimed at enhancing AI accuracy on a global scale.
Funding and Collaboration:
– The study was supported by various institutions and funding bodies aiming to improve eye cancer detection.
– Contributions from contributors are acknowledged, emphasizing collaborative efforts in this field.
Overview of Artificial Intelligence in Ocular Oncology:
– Artificial intelligence (AI) is transforming how retinoblastoma, a pediatric eye cancer, is diagnosed and managed.
– AI techniques such as deep learning are increasingly utilized to analyze medical images and improve diagnostic accuracy.
Deep Learning Models for Tumor Classification:
– Recent studies highlight the use of deep learning models combined with optimization algorithms for ocular tumor classification.
– These models demonstrate significant improvements in identifying retinoblastoma from ocular images.
Global Presentation and Treatment Analysis:
– An international study provided insights into the presentation of retinoblastoma and treatment disparities across different income levels.
– Identifying socio-economic factors is crucial for improving detection and treatment outcomes globally.
International Lag Time in Treatment:
– Research reveals a concerning lag between the onset of retinoblastoma symptoms and the initiation of treatment, averaging across multiple countries.
– This delay can significantly impact patient prognoses and highlights the need for improved awareness and screening.
Data Quality in Deep Learning Applications:
– The effectiveness of deep learning models in medical imaging relies heavily on the quality, type, and volume of data used.
– Continuous advancements in algorithm development are necessary to enhance data utilization for accurate diagnostic outcomes.
Ethnic Variations in Ocular Imaging:
– Studies show ethnic differences in retinal imaging metrics, emphasizing the importance of diversity in biometric studies.
– Research aims to understand how these variations can affect AI outcomes in ocular health assessments.
AI in Retinopathy of Prematurity:
– Recent advances in AI are also shaping the diagnosis and management of retinopathy of prematurity.
– Potential AI-driven solutions are being developed for early detection to mitigate risks associated with this condition.
Future Directions of AI in Ophthalmology:
– The integration of AI in ophthalmology is expected to continue growing, particularly in screening and diagnostic applications.
– Future research aims to further refine these technologies to reduce racial bias and enhance clinical accuracy.
Overview of Methodology:
– The article presents a methodology for processing retinoblastoma images, detailing a three-step process for accurate identification.
– It emphasizes the uniform processing of images across different racial groups.
AI Model Performance:
– The AI model demonstrated effective detection and classification of retinoblastoma, with performance metrics analyzed over 2473 images.
– Key metrics include accuracy, sensitivity, specificity, positive predictive value, and negative predictive value.
Image Distribution:
– A distribution table illustrates the number of images and eyes used for training, validation, and testing the AI model.
– Deep learning models for optic disc and tumor classification showed considerable training efficiency across varying datasets.
Sensitivity Analysis:
– The AI model’s sensitivity varied across different International Classification of Retinoblastoma (ICRB) groups and races.
– Notably, the sensitivity was highest in the normal cases and decreased in advanced ICRB groups.
Demographics of Study Cohort:
– The study included a diverse cohort, encompassing different racial backgrounds such as White, African American, Hispanic, and Asian Indian.
– Such demographic representation allows for a more comprehensive analysis of the AI model’s effectiveness.
Validation of AI Tools:
– Validation of AI methods involved comparison against traditional diagnostic techniques to assess performance.
– Results highlighted that AI tools could complement existing methods in ocular oncology.
Future Implications:
– The findings suggest a potential for AI in improving early detection and treatment decisions for retinoblastoma.
– Further research is required to refine the model and validate its performance across broader populations.
Publication and Data Accessibility:
– The article is published under the Creative Commons Attribution license, ensuring accessibility for further research.
– Data and methodologies presented are crucial for enhancing AI applications in ocular oncology.
Study Overview:
– The study explores the role of AI and machine learning in ocular oncology, specifically retinoblastoma.
– It presents findings from a multiracial cohort, emphasizing the diversity of the sample.
Research Significance:
– Highlights the potential of artificial intelligence to improve diagnostic and treatment outcomes.
– Demonstrates the importance of inclusive research in developing effective healthcare solutions.
Publication Details:
– Published in the journal Cancers, 2024, volume 16, article 3516.
– Available online with a DOI link for access to the full text.
https://doi.org/10.3390/cancers16203516
–Rashmi Kumari
