Your task is to perform an image segmentation and detect whether there is a crack and where without using the images' mask during the training - you can only use binary labels for the whole images (crack or not). Then it's up to your creativity how to determine which pixels are responsible for the positive prediction. This version of the project is the hardest and therefore can be awarded with a grade of 5.0 if done properly.
Besides high accuracy, it is also important to perform all important steps of the neural network creation and training like data preprocessing, architecture selection, and results analysis. 

Although usually domain-specific processing is a nice thing to include in the pipeline in this case it's not allowed. Your solution shall be designed in such a way that it can be directly reused on a different problem of the same type e.g. detect cancer cells in medical images knowing only which images are from a healthy patient and which are not.

Use the following dataset https://www.kaggle.com/datasets/lakshaymiddha/crack-segmentation-dataset. Only the train directory is supposed to be used during the training and from masks, only binary information regarding the whole image can be used.