Enhancing Heart Disease Detection in Echocardiogram Images Using Optimized EfficientNetB3 Architecture

Abstract

Despite the progress in treatment options available for heart disease, it remains one of the most common causes of death worldwide emphasizing a necessity to identify informative and effective diagnostic protocols. This paper proposed an efficient deep learning model, EfficientNetB3, to detect Critical Congenital Heart Disease (CCHD) from echocardiogram images and provides accurate predictions. EfficientNetB3 is further fine-tuned specifically for the heart disease diagnosis in echocardiogram images using advanced techniques like dropout, momentum, and batch normalization to improve overall accuracy of our model. To improve our model architecture and hyperparameters, we carried out a set of experiments. The first experiment conducted was of a baseline model having the pre-training weight, getting the test accuracy as 91.80%. Later improvements, such as the dropout and momentum addition increased this accuracy to 94.14% Finally, the best model configuration with fine-tuned dropout rates and dense layer tweaks yielded a test accuracy of 94.53% Accuracy metrics, confusion matrices and F1 scores were used to assess the performance of the model which outperformed current implementations. This research suggest that the optimized EfficientNetB3 model can be powerful to classify different heart diseases with high accuracy, which has a bright outlook for early and accurate diagnosis. This work is a significant addition to the field of computer-aided diagnostic systems for cardiology and promises to improve clinical practice by offering robust, timely diagnostics assistance in cases where expert echocardiogram interpretation resources may not be readily available. Moving forward, additional deep learning architectures, and multi-modal data will be integrated to enhance diagnostic proficiency.