They don't publish fake results. You may not run with the same hyper-parameters for every method. You may download the code related to each paper from a trusted journal or conference and run it in the same hardware condition, then show the results.

https://conferences.miccai.org/2023/papers/191-Paper1896.html

https://www.nature.com/articles/s41598-019-48995-4

Fake results are never published in journals (In some cases, certain journal websites may be fake and their their documents are available). However, certain details are often omitted or only vaguely presented in research papers when researchers conduct experiments based on specific datasets and models. These undisclosed variables may include precise preprocessing procedures, hyperparameter adjustments such as learning rate modifications, and L1 and L2 regularization strategies. It's important to note that researchers actively strive to find optimal solutions before reaching their expected accuracy levels.

Before training the deep learning model, it is essential to split the dataset using proper split scripts into train, validation, and test sets (There can be a different range of amounts of split data that give the optimal solution). If you observe that the accuracy is not good enough, then you should consider improving the network architecture. If the same dataset or paper showed good accuracy with a particular network architecture, but you are unable to achieve similar results, you may consult with the authors to understand how they attained optimal accuracy. This can be done via email or another platform where they are accessible.

You can also follow books, reference papers, or online articles that precisely mention the kinds of aspects that can provide you with optimal solutions to your current problems. Some recommended resources include: