Comparative Analysis of Classical Machine Learning and Deep Learning Models for Robust Brain Tumor Detection: A Benchmarking Study

Paper: Comparative Analysis of Classical Machine Learning and Deep Learning Models for Robust Brain Tumor Detection: A Benchmarking Study. Code details: Please see the GitHub repository for code and usage instructions: https://github.com/yangzi334/BrainTumorClassification Uploaded Contents in Zenodo: - Primary The primary raw dataset (Brain tumor data.rar) Cleaned primary raw dataset (brain tumor binary data.zip) After splitting 'cleaned primary raw dataset' with ratio 0.7, 0.15, 0.15, 'train.zip', 'validataion.zip', 'test.zip' were generated. - Cross-Dataset Cross-domain raw data (brain tumor cross-domain data.zip) Cleaned Cross-domain (brain tumor cross-domain binary data.zip) Note: In the data preprocessing step, before we generated 'brain tumor external data.zip' , we first used phash alogrithm to compare total images from cross-domain source data (brain tumor external raw.zip) against training images from the primary data (train_folder.zip), any visually identical or nearly identical images will be removed from the cross-dataset. The details can be found in the 'check image leakage.ipynb' in the GitHub. Training models: SVN_HOG_new.zip best_svm_hog_model.pkl — Final support vector machine (SVM) model trained with HOG features. ResNet18.zip best_resnet18_overall.pth — Final fine-tuned resnet18 model weights with best overall validation performance. best_resnet18_run1.pth — Best checkpoint for resnet18 from run 1. best_resnet18_run2.pth — Best checkpoint for resnet18 from run 2. best_resnet18_run3.pth — Best checkpoint for resnet18 from run 3. Vit_B_16_new.zip best_vit_overall.pth — Final fine-tuned ViT-B/16 model weights with best overall validation performance. best_vit_run1.pth — Best checkpoint for ViT-B/16 from run 1. best_vit_run2.pth — Best checkpoint for ViT-B/16 from run 2. best_vit_run3.pth — Best checkpoint for ViT-B/16 from run 3. SimCLR.zip best_simclr_overall.pth — Final fine-tuned simclr model weights with best overall validation performance. best_simclr_run1.pth — Best checkpoint for simclr from run 1. best_simclr_run2.pth — Best checkpoint for simclr from run 2. best_simclr_run3.pth — Best checkpoint for simclr from run 3. best_simclr_encoder_only.pth — SimCLR encoder weights only (trained with contrastive learning, no classifier head). Can be used for downstream tasks