| Abstract: |
Medical image analysis has undergone a transformative shift with the integration of deep learning and modern computer vision algorithms. This study critically examines the performance of convolutional neural network (CNN)-based architectures including ResNet-50, VGG-16, InceptionV3, U-Net, and DenseNet applied across diverse medical imaging modalities such as MRI, CT scan, chest X-ray, and fundus imaging. The primary objective is to compare model accuracy, sensitivity, and specificity across multiple diagnostic tasks including tumor segmentation, pneumonia detection, diabetic retinopathy, and COVID-19 classification. A systematic review methodology was adopted, analysing 50 high-quality empirical studies published between 2016 and 2022, sourced from Google Scholar and PubMed. The central hypothesis posits that deep CNN architectures outperform classical machine learning approaches in medical image classification and segmentation tasks. Results confirm that ResNet-50 achieves up to 96.2% classification accuracy on chest X-rays, while U-Net variants record Dice coefficients exceeding 0.93 on brain MRI datasets. Discussion highlights that transfer learning significantly reduces the training data requirement. The paper concludes that intelligent computer vision approaches are clinically viable, though broader deployment demands standardised datasets and regulatory alignment within India's digital health ecosystem. |
