SEGMENTATION OF RETINAL VESSELS USING SWITCHABLE NORMALIZATION WITH CROSS ENTROPY

Abstract

. The project presents a cutting-edge technique designed to significantly improve the precision in identifying vessels within CT scan imagery. At its core, this method introduces an automated system for the adjustment of pivotal parameters, notably the distortion probability and block size. These adjustments are made in real-time, based on the detection of overfitting signals, allowing the algorithm to maintain a perfect balance that safeguards against both overfitting and underfitting scenarios. This ensures that the model is neither too complex for the data it trains on nor too simplistic to capture the essential patterns. In an effort to further refine the model's capability to classify vessels accurately, the project integrates two sophisticated loss functions: Dice loss and cross-entropy loss. The synergy of these loss functions is expected to enhance the model's sensitivity and specificity in classification tasks, making it a powerful tool for medical diagnostics. Moving away from the traditional and often laborious manual optimization of these parameters, the project employs an automated, efficient strategy. This strategy is anchored in a performance-driven trial-and-error methodology, meticulously guided by the outcomes of test dataset evaluations. By doing so, it leverages empirical evidence to fine-tune the model's parameters, ensuring optimal performance. The automation of this optimization process marks a significant leap towards enhancing both the accuracy and operational efficiency of vessel classification in CT imaging. It promises to expedite the diagnostic process, making it more reliable and less prone to human error. By streamlining this aspect of medical diagnostics, the project stands to offer substantial improvements in the speed and reliability of patient care, setting a new standard for precision in medical imaging analysis.