TLDR: A new hybrid AI model combines deep learning (U-Net) with traditional segmentation methods (SLIC) to accurately detect and segment choroidal nevi in high-resolution fundus images. This approach significantly improves segmentation accuracy, reduces reliance on large datasets, enhances generalizability across different imaging devices, and is computationally efficient, allowing it to run on standard CPUs rather than requiring powerful GPUs. This advancement is crucial for early diagnosis and monitoring of eye lesions with potential for melanoma transformation.
Choroidal nevi, often described as freckles in the eye, are common benign pigmented lesions. While usually harmless, a small percentage carry a risk of transforming into melanoma, a serious form of cancer. Early and accurate detection of these lesions in fundus images – detailed photographs of the eye’s interior – is crucial for improving patient outcomes and preventing vision loss or worse.
Despite significant advancements in AI-based image analysis, precisely identifying and segmenting choroidal nevi remains a considerable challenge. Clinicians without specialized expertise often find it difficult, and existing datasets frequently suffer from low resolution and inconsistent labeling. Deep learning models, such as the widely used U-Net, have shown promise but are heavily reliant on vast quantities of high-quality, annotated training data, which is scarce in the medical field. Furthermore, U-Net models often struggle with high-resolution images, exhibit reduced accuracy when applied to images from different sources (a problem known as domain shift), and demand substantial computational resources like powerful GPUs.
Addressing these critical limitations, a new research paper titled “Hybrid Approach for Enhancing Lesion Segmentation in Fundus Images” introduces an innovative solution. Authored by Mohammadmahdi Eshragh, Emad A. Mohammed, Behrouz Far, Ezekiel Weis, Carol L Shields, Sandor R Ferenczy, and Trafford Crump, this paper proposes a novel hybrid model that combines the strengths of traditional mathematical/clustering segmentation methods with insights from deep learning U-Net models. You can read the full research paper here.
The core idea behind this hybrid approach is to overcome the weaknesses of each method when used in isolation. Traditional segmentation techniques, like Simple Linear Iterative Clustering (SLIC), are effective at grouping pixels based on color and spatial proximity but typically require manual adjustment of parameters. Deep learning models, while automatic, are computationally intensive and data-hungry.
The proposed hybrid model works by first training a small-sized U-Net model (e.g., on 128×128 pixel images) to identify lesions. This small U-Net then extracts crucial information, such as the approximate lesion diameter, its location, and pixel-wise probabilities of lesion presence. This information is then fed into a traditional segmentation method, like SLIC, to automate its parameter selection. This allows the traditional method to perform precise segmentation on much larger, high-resolution fundus images (up to 3900×3900 pixels) without the need for extensive manual input or powerful GPUs for inference.
The results of this hybrid model are compelling. When tested on 1024×1024 fundus images, the hybrid model achieved a Dice coefficient of 89.7% and an Intersection over Union (IoU) of 80.01%. This significantly outperforms the Attention U-Net model, which achieved only 51.3% and 34.2% respectively on the same image size. Beyond accuracy, the hybrid model demonstrated superior generalizability, maintaining better performance on external datasets from different fundus cameras, a common challenge for purely deep learning-based systems.
One of the most significant advantages of this new approach is its computational efficiency. Unlike U-Net models that demand high-end GPUs for fast predictions, the hybrid model can operate effectively on a standard CPU. This makes it a practical and accessible solution for real-world clinical applications, especially for deployment on embedded systems like fundus cameras, which typically lack dedicated GPUs. This reduces hardware dependencies, making advanced diagnostic tools more widely available and reducing energy consumption.
From a clinical perspective, this study offers meaningful improvements. By enabling more accurate identification and precise segmentation of choroidal nevi, the model can support clinicians in monitoring lesion stability over time and detecting early signs of growth. This precision is vital for identifying lesions at risk of malignant transformation, facilitating earlier referral and intervention, and ultimately improving patient outcomes.
Also Read:
- K-Prism: A Unified AI Model for Versatile Medical Image Segmentation
- Advancing Breast Cancer Diagnosis: A Transformer-Based Approach
The researchers acknowledge that future work will focus on further refining these techniques, including exploring mathematical approaches to enhance segmentation accuracy with minimal data dependency, developing robust methods for tracking lesion growth over time, and investigating ways to integrate the SLIC model directly into the U-Net’s loss function to further optimize performance.
