Enhanced Framework for Breast Cancer Detection in PET Images Using Hybrid Graph Convolutional Bidirectional LSTM and Hyperparameter Optimization

Abstract

Breast Cancer Detection (BCD) through Positron Emission Tomography (PET) images remains a crucial area of study for efficient treatment planning and early diagnosis. In combining innovative techniques for Noise Reduction (NR), segmentation, Feature Extraction (FE), Feature Selection (FS), classification, and hyperparameter (HP)tuning, this research offers a thorough framework for PET scans for BCD. First, a Hyper-Averaging Filter (HAF) is applied to PET images to effectively remove noise and enhance image clarity, ensuring more accurate analysis. Subsequently, the Improved BIRCH algorithm is utilized for segmentation, enabling the delineation of regions of interest within the images. Gray-level zone length matrix (GLZLM) features are taken from the segmented regions to present comprehensive texture information, providing important information on the textural characteristics of breast tissue. In addition, to maximize the effectiveness of ensuing classification tasks, the most useful features from the extracted feature set are selected with the FS method known as the minimum redundancy maximum relevance (mRMR). For breast cancer classification, a hybrid optimized Inspection Boosted Graph Convolutional Bidirectional LSTM (long short-term memory) units (O-IBGC-BiLSTM) model is developed. This innovative architecture improves PET image spatial and temporal feature processing by combining Graph Convolutional Networks (GCN) with Bi-LSTM memory units. To enhance model performance, hyperparameter tuning is performed using the improved sparrow search algorithm (ISSA), optimizing model parameters for improved accuracy and robustness. It uses the QIN-Breast Dataset to assess the suggested framework, demonstrating its effectiveness in accurately detecting Breast Cancer (BC) in PET images. Overall, this study presents a comprehensive and integrated approach for breast cancer detection, in clinical practice, may improve early diagnosis and treatment.