Meeting notes
Similar github
CNN Im Preprocessing
DL Driven Tumor MRI
ResNet50 Transfer Learning
MRI Alzheimers:
- Optimizing ResNet50 performance using stochastic gradient descent on MRI images for Alzheimer’s disease classification
- Mainly using ResNet50 and retraining to apply to MRI alzheimers, and testing influence of different learning rates
- Uses MRi scans from the ADNI, with four labels
- Very good performance, 97.66 test acc
- Data prep: normalization, augmentation, feature extraction, data splitting
- “images undergo preprocessing to ensure they are prepared, augmented, resized, standardized, and suitable for accurate analysis by the models.”
- A fine-tuned convolutional neural network model for accurate Alzheimer’s disease classification
- “Use of transfer learning for MRI-based Alzheimer’s detection: we utilized three pre-trained deep CNNs—AlexNet, GoogLeNet, and MobileNetV2”
- 2 datasets: 5000 img kaggle data set and 382 MRI img OASIS dataset
- Classes imbalenced, so augmentations used (rotation)
- Im not sure how serious it is, but they do CRAZY augmentation (23->460 in one class)
- THEN it seems like they split on the augmented dataset, not the original data. Wouldn’t this be a massive data leakage? (testing the same brain, but at different angles effectively)
- Optimizing Pretrained ResNet Models for Alzheimer’s Disease Detection in MRI Images
- Same thing, uses ADNI MRI to classify 4 class alzheimers
- Retrains ResNets:50/101/152.
- “The models were uniformly trained across 10 epochs with a cross-entropy loss function, a batch size of 16, and a transfer learning rate of 0.0001. During fine-tuning, we systematically froze ten different layers of the model from the top down to enhance performance, adjusting the learning rate to 0.00001 with callbacks for better monitoring”
- First round of transfer learning, then a fine tuning round with freezing model weights
- Results are suspiciously good thought, 99.4. Grain of salt?
- Using a ResNet-18 Network to Detect Features of Alzheimer’s Disease on Functional Magnetic Resonance Imaging: A Failed Replication.
- Take down of another paper, showing how data leakage in brain hold out vs slice hold out train/test splitting can inflate test accuracy