Thesis on Lung and Colon Cancer prediction and detection using deep learning

Abstract

Classification of medical images is essential to the diagnosis of many diseases. This paper presents a thorough analysis of multimodal deep learning fusion methods for problems involving the classification of medical images. In order to improve the precision and resilience of classification tasks, this work investigates pre-trained deep learning architectures (VGG19, ResNet50 and InceptionV3). By fine-tuning these models with a dataset of various medical images that includes categories like lung and colon problems, the research takes advantage of transfer learning. The software, which is TensorFlow-implemented, combines model ensembling methods with picture data generators. It uses a conventional ensemble approach to combine predictions from each individual model as part of a specific fusion strategy. The ensemble model performed well across several classes, achieving a strong accuracy of roughly 97.43% on the validation set. Early stopping criteria were used in the training phase, and the Adam optimizer was used to optimize on categorical cross-entropy loss. In order to reduce overfitting and improve generalization, hyperparameters are fine-tuned using strategies like data augmentation, dropout, batch normalization, and early termination. Confusion matrix analysis further demonstrated the model's ability to correctly categorize the various categories, with high true positive rates and low false positive and false negative rates across all classes. The final ensemble model, which is stored in HDF5 format, provides a solid foundation for accurate image categorization within the dataset.