Learning feature dependencies for precise tumor region detection and segmentation in optical coherence tomography...

Area of Science:

• Ophthalmic imaging analysis
• Medical image segmentation
• Artificial intelligence in healthcare

Background:

• Accurate segmentation of retinal tumors in Optical Coherence Tomography (OCT) is crucial for early diagnosis and treatment planning.
• Conventional deep learning and transformer models face challenges in delineating overlapping and interdependent pixel features in OCT images, limiting segmentation precision.

Purpose of the Study:

• To introduce a novel Dependent Inter-Feature Segmentation Method (DIFSM) for improved localization and segmentation of retinal tumor regions in OCT images.
• To address the limitations of existing methods in handling complex feature interactions within OCT scans.

Main Methods:

• The DIFSM framework integrates advanced image preprocessing, inter-feature dependency analysis, and a Vision Transformer (ViT) architecture.
• It utilizes intensity and gradient analysis to identify overlapping tumor-affected regions and trains the ViT with matched/unmatched inter-feature representations for enhanced contextual learning.
• Experiments were conducted on the OCTID dataset, evaluating performance using Dice coefficient, IoU, precision, sensitivity, specificity, and MSME, with comparisons to state-of-the-art models.

Main Results:

• DIFSM achieved a Dice coefficient of 96.2% and IoU of 94.8%, with high precision (96.8%), sensitivity (96.6%), and specificity (96.7%).
• The model demonstrated a significant improvement in segmentation accuracy (14.39%) and precision (14.11%) compared to existing methods, while reducing MSME by 13.5%.
• DIFSM consistently outperformed benchmarks in detecting tumor regions associated with macular hole and central serous retinopathy, showing robustness to noise.

Conclusions:

• The DIFSM framework effectively models inter-feature dependencies and resolves overlapping pixel ambiguities using a Vision Transformer, overcoming limitations of current OCT segmentation methods.
• Significant improvements in accuracy and error reduction indicate DIFSM's potential as a reliable tool for automated retinal tumor detection in clinical practice.
• DIFSM offers a promising advancement for OCT-based diagnostic systems, aiding ophthalmologists in early disease identification and treatment planning.