Optimal Image Reconstruction and Anomaly Detection in Diffuse Optical Tomography with Hybrid CNN-LSTM

Biomedical approaches play a prominent role in therapeutic and diagnostic fields, significantly enhancing patient healthcare. Image-guided therapies have greatly minimized the risks associated with human anomalies by enhancing the precision of surgical interventions and disease detection. Diffuse Optical Tomography (DOT) is a novel technique used for diagnosing abnormalities in soft tissues. DOT produces 3D images through the application of physical principles and sophisticated mathematical methods, improving both image resolution and positional accuracy. The effectiveness of optical imaging relies heavily on the quality of image reconstruction. Consequently, the overall quality of DOT images is greatly influenced by the accuracy of the reconstruction process. Recent advancements have introduced numerous reconstruction algorithms for DOT systems. Despite these developments, many of these algorithms still struggle with accurately predicting anomalies and ensuring precise reconstruction. Considering this issue, the current study presents a novel method of hybrid deep learning approach using CNN-LSTM. This method is designed to detect anomalies in the DOT dataset and classify images as either normal or infected. The pre-processing is performed with the Pixel Repetition Method (PRM), feature selection is practiced with a Genetic algorithm (GA) and the classification is carried out with the assistance of a hybrid CNN-LSTM approach. This integration is intended to enhance image reconstruction and improve anomaly detection accuracy. Additionally, simulation results are provided for the MNIST (Modified National Institute of Standards and Technology) dataset. A comparative analysis is then conducted to evaluate the effectiveness of the proposed approach against conventional methods. To evaluate the model's efficiency, performance metrics such as SSIM (Structural Similarity Index Measure), SNR (Signal-to-Noise Ratio), and PSNR (Peak Signal-to-Noise Ratio) are considered. The achieved values are an SSIM of 0.9856, an execution time of 90 ms, an SNR of 8.256, and a PSNR of 32.15. These results are compared with conventional methods and demonstrate the effectiveness of the proposed system.