LIGHTWEIGHT DEEP LEARNING MODELS FOR ACCURATE CHEST X-RAY DISEASE CLASSIFICATION

Abstract

This Thesis try to create a DL frameworke for automatic two class classification of chest Xray images (CXRI). These classe is Normal and Pneumonia. As per WHO, Pneumonia continues to be a prominent respiratory illness globally [9]. Due to radiol ogists’ technical limits and workload, it has been observed with considerable delays in diagnosis for vulnerable populations, including children, the elderly, and immunocom promised patients, as well as the great majority of the general population. Automatic screening devices can help physicians with their work and ease some of their workload. I introduce a pipeline that enables reproducibly building and comparing a multitude of model families through its unified access to a single data pipeline, evaluation engine and logging standard. It utilizes the following architectures: (i) a custom Convolutional Neural Network (CNN), trained from scratch; (ii) MobileNetV2, chosen as a candidate architecture for lightweight deployment; (iii) EfficientNetB0, offering an acceptable computational cost in the context of transferlearning and (iv) ResNet50, usedasahigher capacity residual network. Themethodologystartsfromthedata, followingastructured pre-processing protocol including: file integrity checking prior to pixel extraction, stratified training/validation/testing sets to maintain initial proportions between classes, pixel value normalization and finally training-time augmentation to mitigate impacts from variability in acquisition parameters. I include random flipping, rotation, zoom and variation of brightness and contrast at the time of training for all images to better generalise to data of varying acquired quality. Sample weighted loss was chosen instead of class duplication for the case of imbalance to prevent unnatural upscaling that could lead to overfit on certain examples of overrepresented class and also to train the model to adhere to natural distribution of the input data during classification. Evaluation metrics vary beyond Accuracy and consider threshold sensitivity. Altho, MOBILENetV2withlessparametersoutperformedscratchtrainednetwork, it also maintained performance and is most desirable for deployment-constrained use cases. Interpretability of learned features in all four models was also investigated using GradCAM visibilization which indicated that models learnt meaningful pulmonary features in a relevant anatomical region. The entire pipeline can be extended to new architectures, augmentation procedures, loss functions, and evaluation cirteria..