Deep learning with PyTorch | Medical Imaging Competitions

Read, split, and display images for a COVID vs non-COVID binary classification using PyTorch, applying augmentation and normalization in a data pipeline with transforms and a dataloader.

Create a custom PyTorch dataset with albumentations to perform multiclass classification on grayscale chest x-ray images across three classes such as covid, normal, and pneumonia.

Leverage a pretrained x-ray vision model for a multiclass task with three classes—covid, pneumonia, and not among the dataset—by building a custom Lightning module and using cross-entropy loss.

apply cross-validation, prepare train and validation splits with a subset random sampler, configure a reduce-on-plateau scheduler and callbacks for model checkpointing, learning-rate monitoring, and early stopping to save the model.

Learn to perform multilevel classification on a two-image medical dataset by processing left and right eye images, combining their features to classify eight labels, and evaluating with the provided script.

Create a PyTorch Lightning module for medical imaging competitions, implementing training and validation steps, using left-right image inputs, replacing BCELoss with focal loss, and integrating F1 score and AUC.

Train and validate the model by evaluating the ODIHR metric on the validation dataset, troubleshoot training and validation steps, and manage memory during parallel model training.

Create the test submission file by loading test data, running the model in evaluation mode, applying a 0.5 threshold to predictions, and saving to submissions.csv.

Frame COVID-19 infection percentage estimation as a regression problem, using mean absolute error and root mean squared error to evaluate predictions. Compare classification versus regression and reference the evaluation criteria.

Set up the dataset from csv and image files, train a fivefold group-based cross-validated PyTorch Lightning model using L1 loss and data augmentation, and prepare a submission zip.

Set up an end-to-end video classification workflow, addressing video lengths with truncation and segments, using the Facebook watch video library, and organizing data with a dataframe for training and validation.

Create a video data loader by building an augmentation pipeline with clip samplers and transforms, applying uniform temporal subsampling, normalization, and cropping to video keys from a Kaggle dataset.

Evaluate a PyTorch Lightning model for medical imaging competitions by validating and testing with a classification report, GPU acceleration, and TensorBoard logs, achieving about 97% validation and 96% test accuracy.