Andrew Y's Senior Project Blog

Project Title: Computer aided diagnosis using image segmentation and stable diffusion for higher accuracy than ground zero and standard diagnosis. BASIS Advisor: Natasha Proctor Internship Location: ASU Onsite Mentor: Jianming Liang

Project Abstract

Machine learning (ML) models interpret and diagnose medical images, specifically chest X-rays. They can detect respiratory diseases, locate abnormalities, and provide specific information through image segmentation. Different ML models use Image Classification, Object Detection, and Segmentation to identify chest diseases, with newer architectures such as CNNs and vision transformers producing precise and efficient models for medical diagnosis. The early 2010s used Neural Networks for image classification, but today, more complex CNN architectures such as Vision transformer models, data augmentation, unsupervised learning models, and transfer learning are used to effectively train models for medical diagnosis, particularly for CXRs (Chest X-rays). One specific type of ML model, is called Convolutional neural networks, or CNNs for short. In recent years, Top-performing models for examining NIH chest X-ray datasets include ResNet-101, VGG-16, VGG-19, and SqueezeNet, with an accuracy score ranging between 90.7% and 94.3%. A more state-of-the-art model, called a diffusion model, is used to learn a lot of detail and images through denoising and noising the image, allowing the model to evaluate the chest x-ray images pixel-by-pixel to predict the specific category (disease) it belongs to. What makes them interesting is its ability to provde a more complex understanding of Chest X-ray diseases with the generation of more realistic and RGB images. My method combines RoentGen model (stable diffusion model) and DenseNet-121 Classfiier model to diagnose atelectasis, cardiomegaly, consolidation, Edema, P. Effusion using RoentGen model feature extraction and Convolution Neural Network. This combination model is evaluated against vision autoencoder models, transfer learning models, and traditional diagnosis (radiologist interperations, clinical assessments, and labroartory tests (ground truth)) based on quantiative metrics and qualitaitve evaluation to identify specific features of diseases. With the help of CheXpert Validation Dataset, NIH Chest X-ray 14 dataset, and RoentGen chest x-ray datasets, I will be able to evaluate whether this model is significantly more effective than SOTA models and traditional diagnosis used today for future use in treating chest & cardiovascular diseases.