Page 45 - AIH-1-4

Artificial Intelligence in Health Segmentation and classification of DR using CNN

Figure 7. Diabetic retinopathy (DR) detection and grading system interface. Image generated using Canva.com.

display at the top of the image promptly communicates the the dynamic nature of LLMs enables continuous learning
severity of the detected condition with a numerical value and refinement, ensuring adaptability to evolving clinical
(0, 1, 2, 3, or 4). Following segmentation results, users scenarios and enhancing the efficacy of predictive analytics
are prompted to input relevant information for further in health care. Overall, our findings underscore the
analysis, with the data processed by a pre-trained LLM to potential of leveraging LLMs in conjunction with image
recommend a suitable treatment strategy for the identified grading techniques to optimize clinical workflows and
DR class. The interface culminates in the generation of improve patient outcomes in the field of ophthalmology. 32
a downloadable PDF report, consolidating segmented
images, normal and grayscale views, classification 4. Conclusion
results, and the recommended treatment, offering a In this research paper, we address the challenges associated
comprehensive summary of the analysis performed by the with the early detection of DR, a severe complication
system (Figure 7). This integrated approach enhances the of diabetes that can lead to blindness. The escalating
diagnostic capabilities, making the interface a user-friendly prevalence of DR underscores the critical need for accurate
and holistic solution for DR detection and classification. 26
and timely diagnosis. Traditional diagnostic methods often
3.4. LLM results face inefficiencies and disagreements among clinicians,
prompting the development of algorithms, particularly
In the domain of medical diagnosis and treatment deep learning approaches, to enhance DR detection. Our
prediction, the utilization of pre-trained LLMs has proposed approach employs transfer learning, leveraging a
emerged as a promising avenue for enhancing clinical single fundus photograph for automatic DR stage detection.
decision-making processes. Leveraging the flexibility Notably, it achieved a commendable ranking in the APTOS
and adaptability of LLMs, our research explored their 2019 Blindness Detection Competition, emphasizing its
application in predicting tests and treatment strategies effectiveness with a high quadratic weighted kappa score
based on the stage predicted by image grading algorithms. of 0.92546. The research aims to contribute significantly to
Using the output from image grading as input for LLMs, we DR detection methodologies, particularly in the context of
capitalized on the comprehensive information extracted automated systems, addressing the crucial requirement for
from retinal images to inform subsequent medical early diagnosis and intervention.
decisions. The inherent capacity of LLMs to comprehend
and contextualize complex medical data enabled them to The study reviews related work, tracing the evolution
provide nuanced predictions tailored to individual patient from classical computer vision approaches to the rise
profiles. This integration of image grading results with LLM- of deep learning, where CNNs have demonstrated
based prediction models facilitated a holistic approach to prowess in DR classification. Transfer learning with
patient care, allowing for more accurate and personalized CNN architectures is explored, highlighting promising
test recommendations and treatment plans. Furthermore, results achieved by various research teams. The problem

Volume 1 Issue 4 (2024) 39 doi:10.36922/aih.2783