A Hybrid Deep Learning Framework for Multi-Disease Detection in Biomedical Images

Abstract

Biomedical image analysis has become an important tool in the diagnosis and treatment planning for a number of diseases in the early stages. Manual image interpretation of biomedical images, however, is time-consuming, requires expertise, and introduce human errors, particularly if there are multiple different diseases and the visual patterns are similar. Traditional machine learning approaches rely on features handcrafted by experts and are not successful in learning intricate structures in medical images relevant to disease. While deep learning models like Convolutional Neural Networks have recently helped with biomedical image detection and classification, the majority of current methods are primarily applicable for detecting single disease and lack capability of modeling inter-dependencies between the abnormalities in image regions. For this reason, this work introduces a hybrid model using a combination of CNN, GNN and RL based networks for the multi-disease detection task in biomedical images. It is designed as a cascading network where CNN networks are used to extract the deep spatial features from the input medical images, the Graph Neural Network learns the relationship among the disease-relevant regions and the Reinforcement Learning determines the informative features and optimizes the classification decision. Experimental evaluation is carried out using the MedMNIST+ benchmark dataset, where multiple biomedical image categories are embraced, encompassing chest X-ray, dermatoscopic images, retinal OCT, retinal fundus and breast ultrasound images. The accuracy and F1 score achieved by the proposed CNN-GNN-RL framework were found to be 96.3% and 95.4% respectively, which improved significantly compared to the precision of the CNN, ResNet50, DenseNet121, CNN-RL and CNN-GNN models. The results show that the proposed method that integrates spatial feature extraction, graph-based relational learning and reinforcement learning based optimization can achieve a higher classification accuracy, lower misclassification rate and stronger robustness of the model compared with the other three methods. Thus, the suggested framework might be helpful to improve the intelligence computer-aided diagnosis system for the accurate and automated detection and diagnosis of multiple diseases in biomedical images.